Delete main

main/app/app.py

@@ -1,218 +0,0 @@
-import os
-import sys
-import torch
-import shutil
-import librosa
-import logging
-import requests
-import subprocess
-import numpy as np
-import gradio as gr
-import soundfile as sf
-from time import sleep
-from multiprocessing import cpu_count
-
-sys.path.append(os.getcwd())
-from main.app.tabs.inference.inference import inference_tabs
-from main.app.tabs.models.model import model_tabs
-from main.app.tabs.utils.utils import utils_tabs
-
-from main.tools import huggingface
-from main.configs.config import Config
-from main.app.based.utils import *
-
-with gr.Blocks(title="Ultimate RVC Maker ⚡", theme=theme) as app:
-    gr.HTML("<h1 style='text-align: center;'>Ultimate RVC Maker ⚡</h1>")
-    gr.Markdown(
-        f"""
-        If you liked this HF Space you can give me a ❤️
-        
-        Try Ultimate RVC Maker WebUI using Colab [here](https://colab.research.google.com/github/TheNeodev/Notebook/blob/main/RVC-MAKER.ipynb)
-        """
-    )
-    with gr.Tabs():      
-        
-
-        with gr.TabItem("Inference"):
-            inference_tabs()
-        with gr.TabItem("Model Options"):
-            model_tabs()
-        with gr.TabItem(translations["separator_tab"], visible=configs.get("separator_tab", True)):
-            gr.Markdown(f"## {translations['separator_tab']}")
-            with gr.Row(): 
-                gr.Markdown(translations["4_part"])
-            with gr.Row():
-                with gr.Column():
-                    with gr.Group():
-                        with gr.Row(equal_height=True):       
-                            cleaner = gr.Checkbox(label=translations["clear_audio"], value=False, interactive=True, min_width=140)       
-                            backing = gr.Checkbox(label=translations["separator_backing"], value=False, interactive=True, min_width=140)
-                            reverb = gr.Checkbox(label=translations["dereveb_audio"], value=False, interactive=True, min_width=140)
-                            backing_reverb = gr.Checkbox(label=translations["dereveb_backing"], value=False, interactive=False, min_width=140)               
-                            denoise = gr.Checkbox(label=translations["denoise_mdx"], value=False, interactive=False, min_width=140)     
-                        with gr.Row(equal_height=True):
-                            separator_model = gr.Dropdown(label=translations["separator_model"], value=uvr_model[0], choices=uvr_model, interactive=True)
-                            separator_backing_model = gr.Dropdown(label=translations["separator_backing_model"], value="Version-1", choices=["Version-1", "Version-2"], interactive=True, visible=backing.value)
-          
-            with gr.Row():
-                with gr.Column():
-                    with gr.Group():
-                        with gr.Row(equal_height=True):
-                            shifts = gr.Slider(label=translations["shift"], info=translations["shift_info"], minimum=1, maximum=20, value=2, step=1, interactive=True)
-                            segment_size = gr.Slider(label=translations["segments_size"], info=translations["segments_size_info"], minimum=32, maximum=3072, value=256, step=32, interactive=True)
-                        with gr.Row():
-                            mdx_batch_size = gr.Slider(label=translations["batch_size"], info=translations["mdx_batch_size_info"], minimum=1, maximum=64, value=1, step=1, interactive=True, visible=backing.value or reverb.value or separator_model.value in mdx_model)
-                with gr.Column():
-                    with gr.Group():
-                        with gr.Row():
-                            overlap = gr.Radio(label=translations["overlap"], info=translations["overlap_info"], choices=["0.25", "0.5", "0.75", "0.99"], value="0.25", interactive=True)
-                        with gr.Row():
-                            mdx_hop_length = gr.Slider(label="Hop length", info=translations["hop_length_info"], minimum=1, maximum=8192, value=1024, step=1, interactive=True, visible=backing.value or reverb.value or separator_model.value in mdx_model)
-                with gr.Column():
-                    with gr.Row():
-                        clean_strength = gr.Slider(label=translations["clean_strength"], info=translations["clean_strength_info"], minimum=0, maximum=1, value=0.5, step=0.1, interactive=True, visible=cleaner.value)
-                        sample_rate1 = gr.Slider(minimum=8000, maximum=96000, step=1, value=44100, label=translations["sr"], info=translations["sr_info"], interactive=True)
-            input = gr.File(label=translations["drop_audio"], file_types=[".wav", ".mp3", ".flac", ".ogg", ".opus", ".m4a", ".mp4", ".aac", ".alac", ".wma", ".aiff", ".webm", ".ac3"]) 
-            audio_input = gr.Audio(show_download_button=True, interactive=False, label=translations["input_audio"])
-            with gr.Column():
-                with gr.Accordion(translations["use_url"], open=False):
-                    url = gr.Textbox(label=translations["url_audio"], value="", placeholder="https://www.youtube.com/...", scale=6)
-                    download_button = gr.Button(translations["downloads"])
-            with gr.Column():
-                with gr.Accordion(translations["input_output"], open=False):
-                    format = gr.Radio(label=translations["export_format"], info=translations["export_info"], choices=["wav", "mp3", "flac", "ogg", "opus", "m4a", "mp4", "aac", "alac", "wma", "aiff", "webm", "ac3"], value="wav", interactive=True)
-                    input_audio = gr.Dropdown(label=translations["audio_path"], value="", choices=paths_for_files, allow_custom_value=True, interactive=True)
-                    refesh_separator = gr.Button(translations["refesh"])
-                output_separator = gr.Textbox(label=translations["output_folder"], value="audios", placeholder="audios", info=translations["output_folder_info"], interactive=True)
-            separator_button = gr.Button(translations["separator_tab"], variant="primary")
-            with gr.Row():
-                gr.Markdown(translations["output_separator"])
-            with gr.Row():
-                instruments_audio = gr.Audio(show_download_button=True, interactive=False, label=translations["instruments"])
-                original_vocals = gr.Audio(show_download_button=True, interactive=False, label=translations["original_vocal"])
-                main_vocals = gr.Audio(show_download_button=True, interactive=False, label=translations["main_vocal"], visible=backing.value)
-                backing_vocals = gr.Audio(show_download_button=True, interactive=False, label=translations["backing_vocal"], visible=backing.value)
-            with gr.Row():
-                separator_model.change(fn=lambda a, b, c: [visible(a or b or c in mdx_model), visible(a or b or c in mdx_model), valueFalse_interactive(a or b or c in mdx_model), visible(c not in mdx_model)], inputs=[backing, reverb, separator_model], outputs=[mdx_batch_size, mdx_hop_length, denoise, shifts])
-                backing.change(fn=lambda a, b, c: [visible(a or b or c in mdx_model), visible(a or b or c in mdx_model), valueFalse_interactive(a or b or c in mdx_model), visible(a), visible(a), visible(a), valueFalse_interactive(a and b)], inputs=[backing, reverb, separator_model], outputs=[mdx_batch_size, mdx_hop_length, denoise, separator_backing_model, main_vocals, backing_vocals, backing_reverb])
-                reverb.change(fn=lambda a, b, c: [visible(a or b or c in mdx_model), visible(a or b or c in mdx_model), valueFalse_interactive(a or b or c in mdx_model), valueFalse_interactive(a and b)], inputs=[backing, reverb, separator_model], outputs=[mdx_batch_size, mdx_hop_length, denoise, backing_reverb])
-            with gr.Row():
-                input_audio.change(fn=lambda audio: audio if os.path.isfile(audio) else None, inputs=[input_audio], outputs=[audio_input])
-                cleaner.change(fn=visible, inputs=[cleaner], outputs=[clean_strength])
-            with gr.Row():
-                input.upload(fn=lambda audio_in: shutil.move(audio_in.name, os.path.join("audios")), inputs=[input], outputs=[input_audio])
-                refesh_separator.click(fn=change_audios_choices, inputs=[input_audio], outputs=[input_audio])
-            with gr.Row():
-                download_button.click(
-                    fn=download_url, 
-                    inputs=[url], 
-                    outputs=[input_audio, audio_input, url],
-                    api_name='download_url'
-                )
-                separator_button.click(
-                    fn=separator_music, 
-                    inputs=[
-                        input_audio, 
-                        output_separator,
-                        format, 
-                        shifts, 
-                        segment_size, 
-                        overlap, 
-                        cleaner, 
-                        clean_strength, 
-                        denoise, 
-                        separator_model, 
-                        separator_backing_model, 
-                        backing,
-                        reverb, 
-                        backing_reverb,
-                        mdx_hop_length,
-                        mdx_batch_size,
-                        sample_rate1
-                    ],
-                    outputs=[original_vocals, instruments_audio, main_vocals, backing_vocals],
-                    api_name='separator_music'
-                )
-        utils_tabs()     
-        with gr.TabItem(translations["settings"], visible=configs.get("settings_tab", True)):
-            gr.Markdown(translations["settings_markdown"])
-            with gr.Row():
-                gr.Markdown(translations["settings_markdown_2"])
-            with gr.Row():
-                toggle_button = gr.Button(translations["change_light_dark"], variant="secondary", scale=2)
-            with gr.Row():
-                with gr.Column():
-                    language_dropdown = gr.Dropdown(label=translations["lang"], interactive=True, info=translations["lang_restart"], choices=configs.get("support_language", "vi-VN"), value=language)
-                    change_lang = gr.Button(translations["change_lang"], variant="primary", scale=2)
-                with gr.Column():
-                    theme_dropdown = gr.Dropdown(label=translations["theme"], interactive=True, info=translations["theme_restart"], choices=configs.get("themes", theme), value=theme, allow_custom_value=True)
-                    changetheme = gr.Button(translations["theme_button"], variant="primary", scale=2)
-            with gr.Row():
-                with gr.Column():
-                    fp_choice = gr.Radio(choices=["fp16","fp32"], value="fp16" if configs.get("fp16", False) else "fp32", label=translations["precision"], info=translations["precision_info"], interactive=True)
-                    fp_button = gr.Button(translations["update_precision"], variant="secondary", scale=2)
-                with gr.Column():
-                    font_choice = gr.Textbox(label=translations["font"], info=translations["font_info"], value=font, interactive=True)
-                    font_button = gr.Button(translations["change_font"])
-            with gr.Row():
-                with gr.Column():
-                    with gr.Accordion(translations["stop"], open=False):
-                        separate_stop = gr.Button(translations["stop_separate"])
-                        convert_stop = gr.Button(translations["stop_convert"])
-                        create_dataset_stop = gr.Button(translations["stop_create_dataset"])
-                        audioldm2_stop = gr.Button(translations["stop_audioldm2"])
-                        with gr.Accordion(translations["stop_training"], open=False):
-                            model_name_stop = gr.Textbox(label=translations["modelname"], info=translations["training_model_name"], value="", placeholder=translations["modelname"], interactive=True)
-                            preprocess_stop = gr.Button(translations["stop_preprocess"])
-                            extract_stop = gr.Button(translations["stop_extract"])
-                            train_stop = gr.Button(translations["stop_training"])
-            with gr.Row():
-                toggle_button.click(fn=None, js="() => {document.body.classList.toggle('dark')}")
-                fp_button.click(fn=change_fp, inputs=[fp_choice], outputs=[fp_choice])
-            with gr.Row():
-                change_lang.click(fn=change_language, inputs=[language_dropdown], outputs=[])
-                changetheme.click(fn=change_theme, inputs=[theme_dropdown], outputs=[])
-                font_button.click(fn=change_font, inputs=[font_choice], outputs=[])
-            with gr.Row():
-                change_lang.click(fn=None, js="setTimeout(function() {location.reload()}, 15000)", inputs=[], outputs=[])
-                changetheme.click(fn=None, js="setTimeout(function() {location.reload()}, 15000)", inputs=[], outputs=[])
-                font_button.click(fn=None, js="setTimeout(function() {location.reload()}, 15000)", inputs=[], outputs=[])
-            with gr.Row():
-                separate_stop.click(fn=lambda: stop_pid("separate_pid", None, False), inputs=[], outputs=[])
-                convert_stop.click(fn=lambda: stop_pid("convert_pid", None, False), inputs=[], outputs=[])
-                create_dataset_stop.click(fn=lambda: stop_pid("create_dataset_pid", None, False), inputs=[], outputs=[])
-            with gr.Row():
-                preprocess_stop.click(fn=lambda model_name_stop: stop_pid("preprocess_pid", model_name_stop, False), inputs=[model_name_stop], outputs=[])
-                extract_stop.click(fn=lambda model_name_stop: stop_pid("extract_pid", model_name_stop, False), inputs=[model_name_stop], outputs=[])
-                train_stop.click(fn=lambda model_name_stop: stop_pid("train_pid", model_name_stop, True), inputs=[model_name_stop], outputs=[])
-            with gr.Row():
-                audioldm2_stop.click(fn=lambda: stop_pid("audioldm2_pid", None, False), inputs=[], outputs=[])
-                
-    
-    with gr.Row(): 
-        gr.Markdown(translations["terms_of_use"])
-        gr.Markdown(translations["exemption"])
-
-    logger.info(translations["start_app"])
-    logger.info(translations["set_lang"].format(lang=language))
-
-    port = configs.get("app_port", 7860)
-
-    for i in range(configs.get("num_of_restart", 5)):
-        try:
-            app.queue().launch(
-                favicon_path=os.path.join("assets", "ico.png"), 
-                server_name=configs.get("server_name", "0.0.0.0"), 
-                server_port=port, 
-                show_error=configs.get("app_show_error", False), 
-                inbrowser="--open" in sys.argv, 
-                share="--share" in sys.argv, 
-                allowed_paths=allow_disk
-            )
-            break
-        except OSError:
-            logger.debug(translations["port"].format(port=port))
-            port -= 1
-        except Exception as e:
-            logger.error(translations["error_occurred"].format(e=e))
-            sys.exit(1)

main/app/based/utils.py

@@ -1,1534 +0,0 @@
-import os
-import re
-import ssl
-import sys
-import json
-import torch
-import codecs
-import shutil
-import asyncio
-import librosa
-import logging
-import datetime
-import platform
-import requests
-import warnings
-import threading
-import subprocess
-import logging.handlers
-
-import numpy as np
-import gradio as gr
-import pandas as pd
-import soundfile as sf
-
-from time import sleep
-from multiprocessing import cpu_count
-
-sys.path.append(os.getcwd())
-
-from main.tools import huggingface
-from main.configs.config import Config
-
-ssl._create_default_https_context = ssl._create_unverified_context
-logger = logging.getLogger(__name__)
-logger.propagate = False
-
-if logger.hasHandlers(): logger.handlers.clear()
-else:
-    console_handler = logging.StreamHandler()
-    console_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
-    console_handler.setFormatter(console_formatter)
-    console_handler.setLevel(logging.INFO)
-    file_handler = logging.handlers.RotatingFileHandler(os.path.join("assets", "logs", "app.log"), maxBytes=5*1024*1024, backupCount=3, encoding='utf-8')
-    file_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
-    file_handler.setFormatter(file_formatter)
-    file_handler.setLevel(logging.DEBUG)
-    logger.addHandler(console_handler)
-    logger.addHandler(file_handler)
-    logger.setLevel(logging.DEBUG)
-
-warnings.filterwarnings("ignore")
-for l in ["httpx", "gradio", "uvicorn", "httpcore", "urllib3"]:
-    logging.getLogger(l).setLevel(logging.ERROR)
-
-config = Config()
-python = sys.executable
-translations = config.translations 
-configs_json = os.path.join("main", "configs", "config.json")
-configs = json.load(open(configs_json, "r"))
-
-os.environ["TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD"] = "1"
-os.environ["TORCH_FORCE_WEIGHTS_ONLY_LOAD"] = "0"
-
-if config.device in ["cpu", "mps"]  and configs.get("fp16", False):
-    logger.warning(translations["fp16_not_support"])
-    configs["fp16"] = config.is_half = False
-    with open(configs_json, "w") as f:
-        json.dump(configs, f, indent=4)
-
-models, model_options = {}, {}
-
-method_f0 = ["mangio-crepe-full", "crepe-full", "fcpe", "rmvpe", "harvest", "pyin"]
-method_f0_full = ["pm", "dio", "mangio-crepe-tiny", "mangio-crepe-small", "mangio-crepe-medium", "mangio-crepe-large", "mangio-crepe-full", "crepe-tiny", "crepe-small", "crepe-medium", "crepe-large", "crepe-full", "fcpe", "fcpe-legacy", "rmvpe", "rmvpe-legacy", "harvest", "yin", "pyin", "swipe"]
-
-embedders_mode = ["fairseq", "onnx", "transformers", "spin"]
-embedders_model = ["contentvec_base", "hubert_base", "japanese_hubert_base", "korean_hubert_base", "chinese_hubert_base", "portuguese_hubert_base", "custom"]
-
-paths_for_files = sorted([os.path.abspath(os.path.join(root, f)) for root, _, files in os.walk("audios") for f in files if os.path.splitext(f)[1].lower() in (".wav", ".mp3", ".flac", ".ogg", ".opus", ".m4a", ".mp4", ".aac", ".alac", ".wma", ".aiff", ".webm", ".ac3")])
-model_name, index_path, delete_index = sorted(list(model for model in os.listdir(os.path.join("assets", "weights")) if model.endswith((".pth", ".onnx")) and not model.startswith("G_") and not model.startswith("D_"))), sorted([os.path.join(root, name) for root, _, files in os.walk(os.path.join("assets", "logs"), topdown=False) for name in files if name.endswith(".index") and "trained" not in name]), sorted([os.path.join("assets", "logs", f) for f in os.listdir(os.path.join("assets", "logs")) if "mute" not in f and os.path.isdir(os.path.join("assets", "logs", f))])
-pretrainedD, pretrainedG, Allpretrained = ([model for model in os.listdir(os.path.join("assets", "models", "pretrained_custom")) if model.endswith(".pth") and "D" in model], [model for model in os.listdir(os.path.join("assets", "models", "pretrained_custom")) if model.endswith(".pth") and "G" in model], [os.path.join("assets", "models", path, model) for path in ["pretrained_v1", "pretrained_v2", "pretrained_custom"] for model in os.listdir(os.path.join("assets", "models", path)) if model.endswith(".pth") and ("D" in model or "G" in model)])
-
-separate_model = sorted([os.path.join("assets", "models", "uvr5", models) for models in os.listdir(os.path.join("assets", "models", "uvr5")) if models.endswith((".th", ".yaml", ".onnx"))])
-presets_file = sorted(list(f for f in os.listdir(os.path.join("assets", "presets")) if f.endswith(".json")))
-f0_file = sorted([os.path.abspath(os.path.join(root, f)) for root, _, files in os.walk(os.path.join("assets", "f0")) for f in files if f.endswith(".txt")])
-
-language, theme, edgetts, google_tts_voice, mdx_model, uvr_model, font = configs.get("language", "vi-VN"), configs.get("theme", "NoCrypt/miku"), configs.get("edge_tts", ["vi-VN-HoaiMyNeural", "vi-VN-NamMinhNeural"]), configs.get("google_tts_voice", ["vi", "en"]), configs.get("mdx_model", "MDXNET_Main"), (configs.get("demucs_model", "HD_MMI") + configs.get("mdx_model", "MDXNET_Main")), configs.get("font", "https://fonts.googleapis.com/css2?family=Courgette&display=swap")
-
-csv_path = os.path.join("assets", "spreadsheet.csv")
-logger.info(config.device)
-
-if "--allow_all_disk" in sys.argv:
-    import win32api
-
-    allow_disk = win32api.GetLogicalDriveStrings().split('\x00')[:-1]
-else: allow_disk = []
-
-if language == "vi-VN": 
-    import gradio.strings
-    gradio.strings.en = {"RUNNING_LOCALLY": "* Chạy trên liên kết nội bộ:  {}://{}:{}", "RUNNING_LOCALLY_SSR": "* Chạy trên liên kết nội bộ:  {}://{}:{}, với SSR ⚡ (thử nghiệm, để tắt hãy dùng `ssr=False` trong `launch()`)", "SHARE_LINK_DISPLAY": "* Chạy trên liên kết công khai: {}", "COULD_NOT_GET_SHARE_LINK": "\nKhông thể tạo liên kết công khai. Vui lòng kiểm tra kết nối mạng của bạn hoặc trang trạng thái của chúng tôi: https://status.gradio.app.", "COULD_NOT_GET_SHARE_LINK_MISSING_FILE": "\nKhông thể tạo liên kết công khai. Thiếu tập tin: {}. \n\nVui lòng kiểm tra kết nối internet của bạn. Điều này có thể xảy ra nếu phần mềm chống vi-rút của bạn chặn việc tải xuống tệp này. Bạn có thể cài đặt thủ công bằng cách làm theo các bước sau: \n\n1. Tải xuống tệp này: {}\n2. Đổi tên tệp đã tải xuống thành: {}\n3. Di chuyển tệp đến vị trí này: {}", "COLAB_NO_LOCAL": "Không thể hiển thị giao diện nội bộ trên google colab, liên kết công khai đã được tạo.", "PUBLIC_SHARE_TRUE": "\nĐể tạo một liên kết công khai, hãy đặt `share=True` trong `launch()`.", "MODEL_PUBLICLY_AVAILABLE_URL": "Mô hình được cung cấp công khai tại: {} (có thể mất tới một phút để sử dụng được liên kết)", "GENERATING_PUBLIC_LINK": "Đang tạo liên kết công khai (có thể mất vài giây...):", "BETA_INVITE": "\nCảm ơn bạn đã là người dùng Gradio! Nếu bạn có thắc mắc hoặc phản hồi, vui lòng tham gia máy chủ Discord của chúng tôi và trò chuyện với chúng tôi: https://discord.gg/feTf9x3ZSB", "COLAB_DEBUG_TRUE": "Đã phát hiện thấy sổ tay Colab. Ô này sẽ chạy vô thời hạn để bạn có thể xem lỗi và nhật ký. " "Để tắt, hãy đặt debug=False trong launch().", "COLAB_DEBUG_FALSE": "Đã phát hiện thấy sổ tay Colab. Để hiển thị lỗi trong sổ ghi chép colab, hãy đặt debug=True trong launch()", "COLAB_WARNING": "Lưu ý: việc mở Chrome Inspector có thể làm hỏng bản demo trong sổ tay Colab.", "SHARE_LINK_MESSAGE": "\nLiên kết công khai sẽ hết hạn sau 72 giờ. Để nâng cấp GPU và lưu trữ vĩnh viễn miễn phí, hãy chạy `gradio deploy` từ terminal trong thư mục làm việc để triển khai lên huggingface (https://huggingface.co/spaces)", "INLINE_DISPLAY_BELOW": "Đang tải giao diện bên dưới...", "COULD_NOT_GET_SHARE_LINK_CHECKSUM": "\nKhông thể tạo liên kết công khai. Tổng kiểm tra không khớp cho tập tin: {}."}
-
-if os.path.exists(csv_path): cached_data = pd.read_csv(csv_path) 
-else:
-    cached_data = pd.read_csv(codecs.decode("uggcf://qbpf.tbbtyr.pbz/fcernqfurrgf/q/1gNHnDeRULtEfz1Yieaw14USUQjWJy0Oq9k0DrCrjApb/rkcbeg?sbezng=pfi&tvq=1977693859", "rot13"))
-    cached_data.to_csv(csv_path, index=False)
-
-for _, row in cached_data.iterrows():
-    filename = row['Filename']
-    url = None
-
-    for value in row.values:
-        if isinstance(value, str) and "huggingface" in value:
-            url = value
-            break
-
-    if url: models[filename] = url
-
-
-
-def gr_info(message):
-    gr.Info(message, duration=2)
-    logger.info(message)
-
-def gr_warning(message):
-    gr.Warning(message, duration=2)
-    logger.warning(message)
-
-def gr_error(message):
-    gr.Error(message=message, duration=6)
-    logger.error(message)
-
-def get_gpu_info():
-    ngpu = torch.cuda.device_count()
-    gpu_infos = [f"{i}: {torch.cuda.get_device_name(i)} ({int(torch.cuda.get_device_properties(i).total_memory / 1024 / 1024 / 1024 + 0.4)} GB)" for i in range(ngpu) if torch.cuda.is_available() or ngpu != 0]
-    return "\n".join(gpu_infos) if len(gpu_infos) > 0 else translations["no_support_gpu"]
-
-def change_f0_choices(): 
-    f0_file = sorted([os.path.abspath(os.path.join(root, f)) for root, _, files in os.walk(os.path.join("assets", "f0")) for f in files if f.endswith(".txt")])
-    return {"value": f0_file[0] if len(f0_file) >= 1 else "", "choices": f0_file, "__type__": "update"}
-
-def change_audios_choices(input_audio): 
-    audios = sorted([os.path.abspath(os.path.join(root, f)) for root, _, files in os.walk("audios") for f in files if os.path.splitext(f)[1].lower() in (".wav", ".mp3", ".flac", ".ogg", ".opus", ".m4a", ".mp4", ".aac", ".alac", ".wma", ".aiff", ".webm", ".ac3")])
-    return {"value": input_audio if input_audio != "" else (audios[0] if len(audios) >= 1 else ""), "choices": audios, "__type__": "update"}
-
-def change_separate_choices():
-    return [{"choices": sorted([os.path.join("assets", "models", "uvr5", models) for models in os.listdir(os.path.join("assets", "models", "uvr5")) if model.endswith((".th", ".yaml", ".onnx"))]), "__type__": "update"}]
-
-def change_models_choices():
-    model, index = sorted(list(model for model in os.listdir(os.path.join("assets", "weights")) if model.endswith((".pth", ".onnx")) and not model.startswith("G_") and not model.startswith("D_"))), sorted([os.path.join(root, name) for root, _, files in os.walk(os.path.join("assets", "logs"), topdown=False) for name in files if name.endswith(".index") and "trained" not in name])
-    return [{"value": model[0] if len(model) >= 1 else "", "choices": model, "__type__": "update"}, {"value": index[0] if len(index) >= 1 else "", "choices": index, "__type__": "update"}]
-
-def change_allpretrained_choices():
-    return [{"choices": sorted([os.path.join("assets", "models", path, model) for path in ["pretrained_v1", "pretrained_v2", "pretrained_custom"] for model in os.listdir(os.path.join("assets", "models", path)) if model.endswith(".pth") and ("D" in model or "G" in model)]), "__type__": "update"}]
-
-def change_pretrained_choices():
-    return [{"choices": sorted([model for model in os.listdir(os.path.join("assets", "models", "pretrained_custom")) if model.endswith(".pth") and "D" in model]), "__type__": "update"}, {"choices": sorted([model for model in os.listdir(os.path.join("assets", "models", "pretrained_custom")) if model.endswith(".pth") and "G" in model]), "__type__": "update"}]
-
-def change_choices_del():
-    return [{"choices": sorted(list(model for model in os.listdir(os.path.join("assets", "weights")) if model.endswith(".pth") and not model.startswith("G_") and not model.startswith("D_"))), "__type__": "update"}, {"choices": sorted([os.path.join("assets", "logs", f) for f in os.listdir(os.path.join("assets", "logs")) if "mute" not in f and os.path.isdir(os.path.join("assets", "logs", f))]), "__type__": "update"}]
-
-def change_preset_choices():
-    return {"value": "", "choices": sorted(list(f for f in os.listdir(os.path.join("assets", "presets")) if f.endswith(".json"))), "__type__": "update"}
-
-def change_tts_voice_choices(google):
-    return {"choices": google_tts_voice if google else edgetts, "value": google_tts_voice[0] if google else edgetts[0], "__type__": "update"}
-
-def change_backing_choices(backing, merge):
-    if backing or merge: return {"value": False, "interactive": False, "__type__": "update"}
-    elif not backing or not merge: return  {"interactive": True, "__type__": "update"}
-    else: gr_warning(translations["option_not_valid"])
-
-def change_download_choices(select):
-    selects = [False]*10
-
-    if select == translations["download_url"]: selects[0] = selects[1] = selects[2] = True
-    elif select == translations["download_from_csv"]:  selects[3] = selects[4] = True
-    elif select == translations["search_models"]: selects[5] = selects[6] = True
-    elif select == translations["upload"]: selects[9] = True
-    else: gr_warning(translations["option_not_valid"])
-
-    return [{"visible": selects[i], "__type__": "update"} for i in range(len(selects))]
-
-def change_download_pretrained_choices(select):
-    selects = [False]*8
-
-    if select == translations["download_url"]: selects[0] = selects[1] = selects[2] = True
-    elif select == translations["list_model"]: selects[3] = selects[4] = selects[5] = True
-    elif select == translations["upload"]: selects[6] = selects[7] = True
-    else: gr_warning(translations["option_not_valid"])
-
-    return [{"visible": selects[i], "__type__": "update"} for i in range(len(selects))]
-
-def get_index(model):
-    model = os.path.basename(model).split("_")[0]
-    return {"value": next((f for f in [os.path.join(root, name) for root, _, files in os.walk(os.path.join("assets", "logs"), topdown=False) for name in files if name.endswith(".index") and "trained" not in name] if model.split(".")[0] in f), ""), "__type__": "update"} if model else None
-
-def index_strength_show(index):
-    return {"visible": index != "" and os.path.exists(index), "value": 0.5, "__type__": "update"}
-
-def hoplength_show(method, hybrid_method=None):
-    show_hop_length_method = ["mangio-crepe-tiny", "mangio-crepe-small", "mangio-crepe-medium", "mangio-crepe-large", "mangio-crepe-full", "fcpe", "fcpe-legacy", "yin", "pyin"]
-
-    if method in show_hop_length_method: visible = True
-    elif method == "hybrid":
-        methods_str = re.search("hybrid\[(.+)\]", hybrid_method)
-        if methods_str: methods = [method.strip() for method in methods_str.group(1).split("+")]
-
-        for i in methods:
-            visible = i in show_hop_length_method
-            if visible: break
-    else: visible = False
-    
-    return {"visible": visible, "__type__": "update"}
-
-def visible(value):
-    return {"visible": value, "__type__": "update"}
-
-def valueFalse_interactive(inp): 
-    return {"value": False, "interactive": inp, "__type__": "update"}
-
-def valueEmpty_visible1(inp1): 
-    return {"value": "", "visible": inp1, "__type__": "update"}
-
-def process_input(file_path):
-    file_contents = ""
-
-    if not file_path.endswith(".srt"):
-        with open(file_path, "r", encoding="utf-8") as file:
-            file_contents = file.read()
-
-    gr_info(translations["upload_success"].format(name=translations["text"]))
-    return file_contents
-
-def fetch_pretrained_data():
-    response = requests.get(codecs.decode("uggcf://uhttvatsnpr.pb/NauC/Ivrganzrfr-EIP-Cebwrpg/erfbyir/znva/wfba/phfgbz_cergenvarq.wfba", "rot13"))
-    response.raise_for_status()
-
-    return response.json()
-
-def update_sample_rate_dropdown(model):
-    data = fetch_pretrained_data()
-    if model != translations["success"]: return {"choices": list(data[model].keys()), "value": list(data[model].keys())[0], "__type__": "update"}
-
-def if_done(done, p):
-    while 1:
-        if p.poll() is None: sleep(0.5)
-        else: break
-
-    done[0] = True
-
-def restart_app():
-    global app
-
-    gr_info(translations["15s"])
-    os.system("cls" if platform.system() == "Windows" else "clear")
-    
-    app.close()
-    subprocess.run([python, os.path.join("main", "app", "app.py")] + sys.argv[1:])
-
-def change_language(lang):
-    configs = json.load(open(configs_json, "r"))
-    configs["language"] = lang
-
-    with open(configs_json, "w") as f:
-        json.dump(configs, f, indent=4)
-
-    restart_app()
-
-def change_theme(theme):
-    with open(configs_json, "r") as f:
-        configs = json.load(f)
-
-    configs["theme"] = theme
-    with open(configs_json, "w") as f:
-        json.dump(configs, f, indent=4)
-
-    restart_app()
-
-def change_font(font):
-    with open(configs_json, "r") as f:
-        configs = json.load(f)
-
-    configs["font"] = font
-    with open(configs_json, "w") as f:
-        json.dump(configs, f, indent=4)
-
-    restart_app()
-
-def zip_file(name, pth, index):
-    pth_path = os.path.join("assets", "weights", pth)
-    if not pth or not os.path.exists(pth_path) or not pth.endswith((".pth", ".onnx")): return gr_warning(translations["provide_file"].format(filename=translations["model"]))
-
-    zip_file_path = os.path.join("assets", "logs", name, name + ".zip")
-    gr_info(translations["start"].format(start=translations["zip"]))
-
-    import zipfile
-    with zipfile.ZipFile(zip_file_path, 'w') as zipf:
-        zipf.write(pth_path, os.path.basename(pth_path))
-        if index: zipf.write(index, os.path.basename(index))
-
-    gr_info(translations["success"])
-    return {"visible": True, "value": zip_file_path, "__type__": "update"}
-
-def fetch_models_data(search):
-    all_table_data = [] 
-    page = 1 
-
-    while 1:
-        try:
-            response = requests.post(url=codecs.decode("uggcf://ibvpr-zbqryf.pbz/srgpu_qngn.cuc", "rot13"), data={"page": page, "search": search})
-
-            if response.status_code == 200:
-                table_data = response.json().get("table", "")
-                if not table_data.strip(): break  
-                all_table_data.append(table_data)
-                page += 1
-            else:
-                logger.debug(f"{translations['code_error']} {response.status_code}")
-                break  
-        except json.JSONDecodeError:
-            logger.debug(translations["json_error"])
-            break
-        except requests.RequestException as e:
-            logger.debug(translations["requests_error"].format(e=e))
-            break
-    return all_table_data
-
-def search_models(name):
-    gr_info(translations["start"].format(start=translations["search"]))
-    tables = fetch_models_data(name)
-
-    if len(tables) == 0:
-        gr_info(translations["not_found"].format(name=name))
-        return [None]*2
-    else:
-        model_options.clear()
-        
-        from bs4 import BeautifulSoup
-
-        for table in tables:
-            for row in BeautifulSoup(table, "html.parser").select("tr"):
-                name_tag, url_tag = row.find("a", {"class": "fs-5"}), row.find("a", {"class": "btn btn-sm fw-bold btn-light ms-0 p-1 ps-2 pe-2"})
-                url = url_tag["href"].replace("https://easyaivoice.com/run?url=", "")
-                if "huggingface" in url:
-                    if name_tag and url_tag: model_options[name_tag.text.replace(".onnx", "").replace(".pth", "").replace(".index", "").replace(".zip", "").replace(" ", "_").replace("(", "").replace(")", "").replace("[", "").replace("]", "").replace(",", "").replace('"', "").replace("'", "").replace("|", "").strip()] = url
-
-        gr_info(translations["found"].format(results=len(model_options)))
-        return [{"value": "", "choices": model_options, "interactive": True, "visible": True, "__type__": "update"}, {"value": translations["downloads"], "visible": True, "__type__": "update"}]
-
-def move_files_from_directory(src_dir, dest_weights, dest_logs, model_name):
-    for root, _, files in os.walk(src_dir):
-        for file in files:
-            file_path = os.path.join(root, file)
-            if file.endswith(".index"):
-                model_log_dir = os.path.join(dest_logs, model_name)
-                os.makedirs(model_log_dir, exist_ok=True)
-
-                filepath = os.path.join(model_log_dir, file.replace(' ', '_').replace('(', '').replace(')', '').replace('[', '').replace(']', '').replace(",", "").replace('"', "").replace("'", "").replace("|", "").strip())
-                if os.path.exists(filepath): os.remove(filepath)
-
-                shutil.move(file_path, filepath)
-            elif file.endswith(".pth") and not file.startswith("D_") and not file.startswith("G_"):
-                pth_path = os.path.join(dest_weights, model_name + ".pth")
-                if os.path.exists(pth_path): os.remove(pth_path)
-
-                shutil.move(file_path, pth_path)
-            elif file.endswith(".onnx") and not file.startswith("D_") and not file.startswith("G_"):
-                pth_path = os.path.join(dest_weights, model_name + ".onnx")
-                if os.path.exists(pth_path): os.remove(pth_path)
-
-                shutil.move(file_path, pth_path)
-
-def download_url(url):
-    import yt_dlp
-
-    if not url: return gr_warning(translations["provide_url"])
-    if not os.path.exists("audios"): os.makedirs("audios", exist_ok=True)
-
-    with warnings.catch_warnings():
-        warnings.filterwarnings("ignore")
-        ydl_opts = {"format": "bestaudio/best", "postprocessors": [{"key": "FFmpegExtractAudio", "preferredcodec": "wav", "preferredquality": "192"}], "quiet": True, "no_warnings": True, "noplaylist": True, "verbose": False}
-
-        gr_info(translations["start"].format(start=translations["download_music"]))
-
-        with yt_dlp.YoutubeDL(ydl_opts) as ydl:
-            audio_output = os.path.join("audios", re.sub(r'\s+', '-', re.sub(r'[^\w\s\u4e00-\u9fff\uac00-\ud7af\u0400-\u04FF\u1100-\u11FF]', '', ydl.extract_info(url, download=False).get('title', 'video')).strip()))
-            if os.path.exists(audio_output): shutil.rmtree(audio_output, ignore_errors=True)
-
-            ydl_opts['outtmpl'] = audio_output
-            
-        with yt_dlp.YoutubeDL(ydl_opts) as ydl: 
-            audio_output = audio_output + ".wav"
-            if os.path.exists(audio_output): os.remove(audio_output)
-            
-            ydl.download([url])
-
-        gr_info(translations["success"])
-        return [audio_output, audio_output, translations["success"]]
-
-def download_model(url=None, model=None):
-    if not url: return gr_warning(translations["provide_url"])
-    if not model: return gr_warning(translations["provide_name_is_save"])
-
-    model = model.replace(".onnx", "").replace(".pth", "").replace(".index", "").replace(".zip", "").replace(" ", "_").replace("(", "").replace(")", "").replace("[", "").replace("]", "").replace(",", "").replace('"', "").replace("'", "").replace("|", "").strip()
-    url = url.replace("/blob/", "/resolve/").replace("?download=true", "").strip()
-
-    download_dir = os.path.join("download_model")
-    weights_dir = os.path.join("assets", "weights")
-    logs_dir = os.path.join("assets", "logs")
-
-    if not os.path.exists(download_dir): os.makedirs(download_dir, exist_ok=True)
-    if not os.path.exists(weights_dir): os.makedirs(weights_dir, exist_ok=True)
-    if not os.path.exists(logs_dir): os.makedirs(logs_dir, exist_ok=True)
-    
-    try:
-        gr_info(translations["start"].format(start=translations["download"]))
-
-        if url.endswith(".pth"): huggingface.HF_download_file(url, os.path.join(weights_dir, f"{model}.pth"))
-        elif url.endswith(".onnx"): huggingface.HF_download_file(url, os.path.join(weights_dir, f"{model}.onnx"))
-        elif url.endswith(".index"):
-            model_log_dir = os.path.join(logs_dir, model)
-            os.makedirs(model_log_dir, exist_ok=True)
-
-            huggingface.HF_download_file(url, os.path.join(model_log_dir, f"{model}.index"))
-        elif url.endswith(".zip"):
-            output_path = huggingface.HF_download_file(url, os.path.join(download_dir, model + ".zip"))
-            shutil.unpack_archive(output_path, download_dir)
-
-            move_files_from_directory(download_dir, weights_dir, logs_dir, model)
-        else:
-            if "drive.google.com" in url or "drive.usercontent.google.com" in url:
-                file_id = None
-
-                from main.tools import gdown
-
-                if "/file/d/" in url: file_id = url.split("/d/")[1].split("/")[0]
-                elif "open?id=" in url: file_id = url.split("open?id=")[1].split("/")[0]
-                elif "/download?id=" in url: file_id = url.split("/download?id=")[1].split("&")[0]
-                
-                if file_id:
-                    file = gdown.gdown_download(id=file_id, output=download_dir)
-                    if file.endswith(".zip"): shutil.unpack_archive(file, download_dir)
-
-                    move_files_from_directory(download_dir, weights_dir, logs_dir, model)
-            elif "mega.nz" in url:
-                from main.tools import meganz
-                
-                meganz.mega_download_url(url, download_dir)
-
-                file_download = next((f for f in os.listdir(download_dir)), None)
-                if file_download.endswith(".zip"): shutil.unpack_archive(os.path.join(download_dir, file_download), download_dir)
-
-                move_files_from_directory(download_dir, weights_dir, logs_dir, model)
-            elif "mediafire.com" in url:
-                from main.tools import mediafire
-
-                file = mediafire.Mediafire_Download(url, download_dir)
-                if file.endswith(".zip"): shutil.unpack_archive(file, download_dir)
-
-                move_files_from_directory(download_dir, weights_dir, logs_dir, model)
-            elif "pixeldrain.com" in url:
-                from main.tools import pixeldrain
-
-                file = pixeldrain.pixeldrain(url, download_dir)
-                if file.endswith(".zip"): shutil.unpack_archive(file, download_dir)
-
-                move_files_from_directory(download_dir, weights_dir, logs_dir, model)
-            else:
-                gr_warning(translations["not_support_url"])
-                return translations["not_support_url"]
-        
-        gr_info(translations["success"])
-        return translations["success"]
-    except Exception as e:
-        gr_error(message=translations["error_occurred"].format(e=e))
-        logger.debug(e)
-        return translations["error_occurred"].format(e=e)
-    finally:
-        shutil.rmtree(download_dir, ignore_errors=True)
-
-def save_drop_model(dropbox):
-    weight_folder = os.path.join("assets", "weights")
-    logs_folder = os.path.join("assets", "logs")
-    save_model_temp = os.path.join("save_model_temp")
-
-    if not os.path.exists(weight_folder): os.makedirs(weight_folder, exist_ok=True)
-    if not os.path.exists(logs_folder): os.makedirs(logs_folder, exist_ok=True)
-    if not os.path.exists(save_model_temp): os.makedirs(save_model_temp, exist_ok=True)
-
-    shutil.move(dropbox, save_model_temp)
-
-    try:
-        file_name = os.path.basename(dropbox)
-
-        if file_name.endswith(".pth") and file_name.endswith(".onnx") and file_name.endswith(".index"): gr_warning(translations["not_model"])
-        else:    
-            if file_name.endswith(".zip"):
-                shutil.unpack_archive(os.path.join(save_model_temp, file_name), save_model_temp)
-                move_files_from_directory(save_model_temp, weight_folder, logs_folder, file_name.replace(".zip", ""))
-            elif file_name.endswith((".pth", ".onnx")): 
-                output_file = os.path.join(weight_folder, file_name)
-                if os.path.exists(output_file): os.remove(output_file)
-                
-                shutil.move(os.path.join(save_model_temp, file_name), output_file)
-            elif file_name.endswith(".index"):
-                def extract_name_model(filename):
-                    match = re.search(r"([A-Za-z]+)(?=_v|\.|$)", filename)
-                    return match.group(1) if match else None
-                
-                model_logs = os.path.join(logs_folder, extract_name_model(file_name))
-                if not os.path.exists(model_logs): os.makedirs(model_logs, exist_ok=True)
-                shutil.move(os.path.join(save_model_temp, file_name), model_logs)
-            else: 
-                gr_warning(translations["unable_analyze_model"])
-                return None
-        
-        gr_info(translations["upload_success"].format(name=translations["model"]))
-        return None
-    except Exception as e:
-        gr_error(message=translations["error_occurred"].format(e=e))
-        logger.debug(e)
-        return None
-    finally:
-        shutil.rmtree(save_model_temp, ignore_errors=True)
-
-def download_pretrained_model(choices, model, sample_rate):
-    pretraineds_custom_path = os.path.join("assets", "models", "pretrained_custom")
-    if choices == translations["list_model"]:
-        paths = fetch_pretrained_data()[model][sample_rate]
-
-        if not os.path.exists(pretraineds_custom_path): os.makedirs(pretraineds_custom_path, exist_ok=True)
-        url = codecs.decode("uggcf://uhttvatsnpr.pb/NauC/Ivrganzrfr-EIP-Cebwrpg/erfbyir/znva/cergenvarq_phfgbz/", "rot13") + paths
-
-        gr_info(translations["download_pretrain"])
-        file = huggingface.HF_download_file(url.replace("/blob/", "/resolve/").replace("?download=true", "").strip(), os.path.join(pretraineds_custom_path, paths))
-
-        if file.endswith(".zip"): 
-            shutil.unpack_archive(file, pretraineds_custom_path)
-            os.remove(file)
-
-        gr_info(translations["success"])
-        return translations["success"]
-    elif choices == translations["download_url"]:
-        if not model: return gr_warning(translations["provide_pretrain"].format(dg="D"))
-        if not sample_rate: return gr_warning(translations["provide_pretrain"].format(dg="G"))
-
-        gr_info(translations["download_pretrain"])
-
-        huggingface.HF_download_file(model.replace("/blob/", "/resolve/").replace("?download=true", "").strip(), pretraineds_custom_path)
-        huggingface.HF_download_file(sample_rate.replace("/blob/", "/resolve/").replace("?download=true", "").strip(), pretraineds_custom_path)
-
-        gr_info(translations["success"])
-        return translations["success"]
-
-def fushion_model_pth(name, pth_1, pth_2, ratio):
-    if not name.endswith(".pth"): name = name + ".pth"
-
-    if not pth_1 or not os.path.exists(pth_1) or not pth_1.endswith(".pth"):
-        gr_warning(translations["provide_file"].format(filename=translations["model"] + " 1"))
-        return [translations["provide_file"].format(filename=translations["model"] + " 1"), None]
-    
-    if not pth_2 or not os.path.exists(pth_2) or not pth_2.endswith(".pth"):
-        gr_warning(translations["provide_file"].format(filename=translations["model"] + " 2"))
-        return [translations["provide_file"].format(filename=translations["model"] + " 2"), None]
-    
-    from collections import OrderedDict
-
-    def extract(ckpt):
-        a = ckpt["model"]
-        opt = OrderedDict()
-        opt["weight"] = {}
-
-        for key in a.keys():
-            if "enc_q" in key: continue
-
-            opt["weight"][key] = a[key]
-
-        return opt
-    
-    try:
-        ckpt1 = torch.load(pth_1, map_location="cpu")
-        ckpt2 = torch.load(pth_2, map_location="cpu")
-
-        if ckpt1["sr"] != ckpt2["sr"]: 
-            gr_warning(translations["sr_not_same"])
-            return [translations["sr_not_same"], None]
-
-        cfg = ckpt1["config"]
-        cfg_f0 = ckpt1["f0"]
-        cfg_version = ckpt1["version"]
-        cfg_sr = ckpt1["sr"]
-
-        vocoder = ckpt1.get("vocoder", "Default")
-
-        ckpt1 = extract(ckpt1) if "model" in ckpt1 else ckpt1["weight"]
-        ckpt2 = extract(ckpt2) if "model" in ckpt2 else ckpt2["weight"]
-
-        if sorted(list(ckpt1.keys())) != sorted(list(ckpt2.keys())): 
-            gr_warning(translations["architectures_not_same"])
-            return [translations["architectures_not_same"], None]
-         
-        gr_info(translations["start"].format(start=translations["fushion_model"]))
-
-        opt = OrderedDict()
-        opt["weight"] = {}
-
-        for key in ckpt1.keys():
-            if key == "emb_g.weight" and ckpt1[key].shape != ckpt2[key].shape:
-                min_shape0 = min(ckpt1[key].shape[0], ckpt2[key].shape[0])
-                opt["weight"][key] = (ratio * (ckpt1[key][:min_shape0].float()) + (1 - ratio) * (ckpt2[key][:min_shape0].float())).half()
-            else: opt["weight"][key] = (ratio * (ckpt1[key].float()) + (1 - ratio) * (ckpt2[key].float())).half()
-
-        opt["config"] = cfg
-        opt["sr"] = cfg_sr
-        opt["f0"] = cfg_f0
-        opt["version"] = cfg_version
-        opt["infos"] = translations["model_fushion_info"].format(name=name, pth_1=pth_1, pth_2=pth_2, ratio=ratio)
-        opt["vocoder"] = vocoder
-
-        output_model = os.path.join("assets", "weights")
-        if not os.path.exists(output_model): os.makedirs(output_model, exist_ok=True)
-
-        torch.save(opt, os.path.join(output_model, name))
-
-        gr_info(translations["success"])
-        return [translations["success"], os.path.join(output_model, name)]
-    except Exception as e:
-        gr_error(message=translations["error_occurred"].format(e=e))
-        logger.debug(e)
-        return [e, None]
-
-def fushion_model(name, path_1, path_2, ratio):
-    if not name:
-        gr_warning(translations["provide_name_is_save"]) 
-        return [translations["provide_name_is_save"], None]
-
-    if path_1.endswith(".pth") and path_2.endswith(".pth"): return fushion_model_pth(name.replace(".onnx", ".pth"), path_1, path_2, ratio)
-    else:
-        gr_warning(translations["format_not_valid"])
-        return [None, None]
-    
-def onnx_export(model_path):
-    from main.library.algorithm.onnx_export import onnx_exporter
-    
-    if not model_path.endswith(".pth"): model_path + ".pth"
-    if not model_path or not os.path.exists(model_path) or not model_path.endswith(".pth"):
-        gr_warning(translations["provide_file"].format(filename=translations["model"]))
-        return [None, translations["provide_file"].format(filename=translations["model"])]
-    
-    try:
-        gr_info(translations["start_onnx_export"])
-        output = onnx_exporter(model_path, model_path.replace(".pth", ".onnx"), is_half=config.is_half, device=config.device)
-
-        gr_info(translations["success"])
-        return [output, translations["success"]]
-    except Exception as e:
-        return [None, e]
-    
-def model_info(path):
-    if not path or not os.path.exists(path) or os.path.isdir(path) or not path.endswith((".pth", ".onnx")): return gr_warning(translations["provide_file"].format(filename=translations["model"]))
-    
-    def prettify_date(date_str):
-        if date_str == translations["not_found_create_time"]: return None
-
-        try:
-            return datetime.datetime.strptime(date_str, "%Y-%m-%dT%H:%M:%S.%f").strftime("%Y-%m-%d %H:%M:%S")
-        except ValueError as e:
-            logger.debug(e)
-            return translations["format_not_valid"]
-    
-    if path.endswith(".pth"): model_data = torch.load(path, map_location=torch.device("cpu"))
-    else:
-        import onnx
-        
-        model = onnx.load(path)
-        model_data = None
-
-        for prop in model.metadata_props:
-            if prop.key == "model_info":
-                model_data = json.loads(prop.value)
-                break
-
-    gr_info(translations["read_info"])
-
-    epochs = model_data.get("epoch", None)
-    if epochs is None: 
-        epochs = model_data.get("info", None)
-        try:
-            epoch = epochs.replace("epoch", "").replace("e", "").isdigit()
-            if epoch and epochs is None: epochs = translations["not_found"].format(name=translations["epoch"])
-        except: 
-            pass
-
-    steps = model_data.get("step", translations["not_found"].format(name=translations["step"]))
-    sr = model_data.get("sr", translations["not_found"].format(name=translations["sr"]))
-    f0 = model_data.get("f0", translations["not_found"].format(name=translations["f0"]))
-    version = model_data.get("version", translations["not_found"].format(name=translations["version"]))
-    creation_date = model_data.get("creation_date", translations["not_found_create_time"])
-    model_hash = model_data.get("model_hash", translations["not_found"].format(name="model_hash"))
-    pitch_guidance = translations["trained_f0"] if f0 else translations["not_f0"]
-    creation_date_str = prettify_date(creation_date) if creation_date else translations["not_found_create_time"]
-    model_name = model_data.get("model_name", translations["unregistered"])
-    model_author = model_data.get("author", translations["not_author"])
-    vocoder = model_data.get("vocoder", "Default")
-
-    gr_info(translations["success"])
-    return translations["model_info"].format(model_name=model_name, model_author=model_author, epochs=epochs, steps=steps, version=version, sr=sr, pitch_guidance=pitch_guidance, model_hash=model_hash, creation_date_str=creation_date_str, vocoder=vocoder)
-
-def audio_effects(input_path, output_path, resample, resample_sr, chorus_depth, chorus_rate, chorus_mix, chorus_delay, chorus_feedback, distortion_drive, reverb_room_size, reverb_damping, reverb_wet_level, reverb_dry_level, reverb_width, reverb_freeze_mode, pitch_shift, delay_seconds, delay_feedback, delay_mix, compressor_threshold, compressor_ratio, compressor_attack_ms, compressor_release_ms, limiter_threshold, limiter_release, gain_db, bitcrush_bit_depth, clipping_threshold, phaser_rate_hz, phaser_depth, phaser_centre_frequency_hz, phaser_feedback, phaser_mix, bass_boost_db, bass_boost_frequency, treble_boost_db, treble_boost_frequency, fade_in_duration, fade_out_duration, export_format, chorus, distortion, reverb, delay, compressor, limiter, gain, bitcrush, clipping, phaser, treble_bass_boost, fade_in_out, audio_combination, audio_combination_input):
-    if not input_path or not os.path.exists(input_path) or os.path.isdir(input_path): 
-        gr_warning(translations["input_not_valid"])
-        return None
-        
-    if not output_path:
-        gr_warning(translations["output_not_valid"])
-        return None
-    
-    if os.path.isdir(output_path): output_path = os.path.join(output_path, f"audio_effects.{export_format}")
-    output_dir = os.path.dirname(output_path) or output_path
-
-    if not os.path.exists(output_dir): os.makedirs(output_dir, exist_ok=True)
-    if os.path.exists(output_path): os.remove(output_path)
-    
-    gr_info(translations["start"].format(start=translations["apply_effect"]))
-    subprocess.run([python, "main/inference/audio_effects.py", "--input_path", input_path, "--output_path", output_path, "--resample", str(resample), "--resample_sr", str(resample_sr), "--chorus_depth", str(chorus_depth), "--chorus_rate", str(chorus_rate), "--chorus_mix", str(chorus_mix), "--chorus_delay", str(chorus_delay), "--chorus_feedback", str(chorus_feedback), "--drive_db", str(distortion_drive), "--reverb_room_size", str(reverb_room_size), "--reverb_damping", str(reverb_damping), "--reverb_wet_level", str(reverb_wet_level), "--reverb_dry_level", str(reverb_dry_level), "--reverb_width", str(reverb_width), "--reverb_freeze_mode", str(reverb_freeze_mode), "--pitch_shift", str(pitch_shift), "--delay_seconds", str(delay_seconds), "--delay_feedback", str(delay_feedback), "--delay_mix", str(delay_mix), "--compressor_threshold", str(compressor_threshold), "--compressor_ratio", str(compressor_ratio), "--compressor_attack_ms", str(compressor_attack_ms), "--compressor_release_ms", str(compressor_release_ms), "--limiter_threshold", str(limiter_threshold), "--limiter_release", str(limiter_release), "--gain_db", str(gain_db), "--bitcrush_bit_depth", str(bitcrush_bit_depth), "--clipping_threshold", str(clipping_threshold), "--phaser_rate_hz", str(phaser_rate_hz), "--phaser_depth", str(phaser_depth), "--phaser_centre_frequency_hz", str(phaser_centre_frequency_hz), "--phaser_feedback", str(phaser_feedback), "--phaser_mix", str(phaser_mix), "--bass_boost_db", str(bass_boost_db), "--bass_boost_frequency", str(bass_boost_frequency), "--treble_boost_db", str(treble_boost_db), "--treble_boost_frequency", str(treble_boost_frequency), "--fade_in_duration", str(fade_in_duration), "--fade_out_duration", str(fade_out_duration), "--export_format", export_format, "--chorus", str(chorus), "--distortion", str(distortion), "--reverb", str(reverb), "--pitchshift", str(pitch_shift != 0), "--delay", str(delay), "--compressor", str(compressor), "--limiter", str(limiter), "--gain", str(gain), "--bitcrush", str(bitcrush), "--clipping", str(clipping), "--phaser", str(phaser), "--treble_bass_boost", str(treble_bass_boost), "--fade_in_out", str(fade_in_out), "--audio_combination", str(audio_combination), "--audio_combination_input", audio_combination_input])
-
-    gr_info(translations["success"])
-    return output_path.replace("wav", export_format)
-
-def synthesize_tts(prompt, voice, speed, output, pitch, google):
-    if not google: 
-        from edge_tts import Communicate
-
-        asyncio.run(Communicate(text=prompt, voice=voice, rate=f"+{speed}%" if speed >= 0 else f"{speed}%", pitch=f"+{pitch}Hz" if pitch >= 0 else f"{pitch}Hz").save(output))
-    else: 
-        response = requests.get(codecs.decode("uggcf://genafyngr.tbbtyr.pbz/genafyngr_ggf", "rot13"), params={"ie": "UTF-8", "q": prompt, "tl": voice, "ttsspeed": speed, "client": "tw-ob"}, headers={"User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/111.0.0.0 Safari/537.36"})
-
-        if response.status_code == 200:
-            with open(output, "wb") as f:
-                f.write(response.content)
-
-            if pitch != 0 or speed != 0:
-                y, sr = librosa.load(output, sr=None)
-
-                if pitch != 0: y = librosa.effects.pitch_shift(y, sr=sr, n_steps=pitch)
-                if speed != 0: y = librosa.effects.time_stretch(y, rate=speed)
-
-                sf.write(file=output, data=y, samplerate=sr, format=os.path.splitext(os.path.basename(output))[-1].lower().replace('.', ''))
-        else: gr_error(f"{response.status_code}, {response.text}")
-
-def time_stretch(y, sr, target_duration):
-    rate = (len(y) / sr) / target_duration
-    if rate != 1.0: y = librosa.effects.time_stretch(y=y.astype(np.float32), rate=rate)
-
-    n_target = int(round(target_duration * sr))
-    return np.pad(y, (0, n_target - len(y))) if len(y) < n_target else y[:n_target]
-
-def pysrttime_to_seconds(t):
-    return (t.hours * 60 + t.minutes) * 60 + t.seconds + t.milliseconds / 1000
-
-def srt_tts(srt_file, out_file, voice, rate = 0, sr = 24000, google = False):
-    import pysrt
-    import tempfile
-
-    subs = pysrt.open(srt_file)
-    if not subs: raise ValueError(translations["srt"])
-
-    final_audio = np.zeros(int(round(pysrttime_to_seconds(subs[-1].end) * sr)), dtype=np.float32)
-
-    with tempfile.TemporaryDirectory() as tempdir:
-        for idx, seg in enumerate(subs):
-            wav_path = os.path.join(tempdir, f"seg_{idx}.wav")
-            synthesize_tts(" ".join(seg.text.splitlines()), voice, 0, wav_path, rate, google)
-
-            audio, file_sr = sf.read(wav_path, dtype=np.float32)
-            if file_sr != sr: audio = np.interp(np.linspace(0, len(audio) - 1, int(len(audio) * sr / file_sr)), np.arange(len(audio)), audio)
-            adjusted = time_stretch(audio, sr, pysrttime_to_seconds(seg.duration))
-
-            start_sample = int(round(pysrttime_to_seconds(seg.start) * sr))
-            end_sample = start_sample + adjusted.shape[0]
-
-            if end_sample > final_audio.shape[0]:
-                adjusted = adjusted[: final_audio.shape[0] - start_sample]
-                end_sample = final_audio.shape[0]
-
-            final_audio[start_sample:end_sample] += adjusted
-
-    sf.write(out_file, final_audio, sr)
-
-def TTS(prompt, voice, speed, output, pitch, google, srt_input):
-    if not srt_input: srt_input = ""
-
-    if not prompt and not srt_input.endswith(".srt"):
-        gr_warning(translations["enter_the_text"])
-        return None
-    
-    if not voice:
-        gr_warning(translations["choose_voice"])
-        return None
-    
-    if not output: 
-        gr_warning(translations["output_not_valid"])
-        return None
-    
-    if os.path.isdir(output): output = os.path.join(output, f"tts.wav")
-    gr_info(translations["convert"].format(name=translations["text"]))
-
-    output_dir = os.path.dirname(output) or output
-    if not os.path.exists(output_dir): os.makedirs(output_dir, exist_ok=True)
-
-    if srt_input.endswith(".srt"): srt_tts(srt_input, output, voice, 0, 24000, google)
-    else: synthesize_tts(prompt, voice, speed, output, pitch, google)
-
-    gr_info(translations["success"])
-    return output
-
-def separator_music(input, output_audio, format, shifts, segments_size, overlap, clean_audio, clean_strength, denoise, separator_model, kara_model, backing, reverb, backing_reverb, hop_length, batch_size, sample_rate):
-    output = os.path.dirname(output_audio) or output_audio
-
-    if not input or not os.path.exists(input) or os.path.isdir(input): 
-        gr_warning(translations["input_not_valid"])
-        return [None]*4
-    
-    if not os.path.exists(output): 
-        gr_warning(translations["output_not_valid"])
-        return [None]*4
-
-    if not os.path.exists(output): os.makedirs(output)
-    gr_info(translations["start"].format(start=translations["separator_music"]))
-
-    subprocess.run([python, "main/inference/separator_music.py", "--input_path", input, "--output_path", output, "--format", format, "--shifts", str(shifts), "--segments_size", str(segments_size), "--overlap", str(overlap), "--mdx_hop_length", str(hop_length), "--mdx_batch_size", str(batch_size), "--clean_audio", str(clean_audio), "--clean_strength", str(clean_strength), "--kara_model", kara_model, "--backing", str(backing), "--mdx_denoise", str(denoise), "--reverb", str(reverb), "--backing_reverb", str(backing_reverb), "--model_name", separator_model, "--sample_rate", str(sample_rate)])
-    gr_info(translations["success"])
-
-    filename, _ = os.path.splitext(os.path.basename(input))
-    output = os.path.join(output, filename)
-
-    return [os.path.join(output, f"Original_Vocals_No_Reverb.{format}") if reverb else os.path.join(output, f"Original_Vocals.{format}"), os.path.join(output, f"Instruments.{format}"), (os.path.join(output, f"Main_Vocals_No_Reverb.{format}") if reverb else os.path.join(output, f"Main_Vocals.{format}") if backing else None), (os.path.join(output, f"Backing_Vocals_No_Reverb.{format}") if backing_reverb else os.path.join(output, f"Backing_Vocals.{format}") if backing else None)] if os.path.isfile(input) else [None]*4
-
-def convert(pitch, filter_radius, index_rate, volume_envelope, protect, hop_length, f0_method, input_path, output_path, pth_path, index_path, f0_autotune, clean_audio, clean_strength, export_format, embedder_model, resample_sr, split_audio, f0_autotune_strength, checkpointing, onnx_f0_mode, embedders_mode, formant_shifting, formant_qfrency, formant_timbre, f0_file):    
-    subprocess.run([python, "main/inference/convert.py", "--pitch", str(pitch), "--filter_radius", str(filter_radius), "--index_rate", str(index_rate), "--volume_envelope", str(volume_envelope), "--protect", str(protect), "--hop_length", str(hop_length), "--f0_method", f0_method, "--input_path", input_path, "--output_path", output_path, "--pth_path", pth_path, "--index_path", index_path if index_path else "", "--f0_autotune", str(f0_autotune), "--clean_audio", str(clean_audio), "--clean_strength", str(clean_strength), "--export_format", export_format, "--embedder_model", embedder_model, "--resample_sr", str(resample_sr), "--split_audio", str(split_audio), "--f0_autotune_strength", str(f0_autotune_strength), "--checkpointing", str(checkpointing), "--f0_onnx", str(onnx_f0_mode), "--embedders_mode", embedders_mode, "--formant_shifting", str(formant_shifting), "--formant_qfrency", str(formant_qfrency), "--formant_timbre", str(formant_timbre), "--f0_file", f0_file])
-
-def convert_audio(clean, autotune, use_audio, use_original, convert_backing, not_merge_backing, merge_instrument, pitch, clean_strength, model, index, index_rate, input, output, format, method, hybrid_method, hop_length, embedders, custom_embedders, resample_sr, filter_radius, volume_envelope, protect, split_audio, f0_autotune_strength, input_audio_name, checkpointing, onnx_f0_mode, formant_shifting, formant_qfrency, formant_timbre, f0_file, embedders_mode):
-    model_path = os.path.join("assets", "weights", model)
-
-    return_none = [None]*6
-    return_none[5] = {"visible": True, "__type__": "update"}
-
-    if not use_audio:
-        if merge_instrument or not_merge_backing or convert_backing or use_original:
-            gr_warning(translations["turn_on_use_audio"])
-            return return_none
-
-    if use_original:
-        if convert_backing:
-            gr_warning(translations["turn_off_convert_backup"])
-            return return_none
-        elif not_merge_backing:
-            gr_warning(translations["turn_off_merge_backup"])
-            return return_none
-
-    if not model or not os.path.exists(model_path) or os.path.isdir(model_path) or not model.endswith((".pth", ".onnx")):
-        gr_warning(translations["provide_file"].format(filename=translations["model"]))
-        return return_none
-
-    f0method, embedder_model = (method if method != "hybrid" else hybrid_method), (embedders if embedders != "custom" else custom_embedders)
-
-    if use_audio:
-        output_audio = os.path.join("audios", input_audio_name)
-
-        from main.library.utils import pydub_convert, pydub_load
-        
-        def get_audio_file(label):
-            matching_files = [f for f in os.listdir(output_audio) if label in f]
-
-            if not matching_files: return translations["notfound"]   
-            return os.path.join(output_audio, matching_files[0])
-
-        output_path = os.path.join(output_audio, f"Convert_Vocals.{format}")
-        output_backing = os.path.join(output_audio, f"Convert_Backing.{format}")
-        output_merge_backup = os.path.join(output_audio, f"Vocals+Backing.{format}")
-        output_merge_instrument = os.path.join(output_audio, f"Vocals+Instruments.{format}")
-
-        if os.path.exists(output_audio): os.makedirs(output_audio, exist_ok=True)
-        if os.path.exists(output_path): os.remove(output_path)
-
-        if use_original:
-            original_vocal = get_audio_file('Original_Vocals_No_Reverb.')
-
-            if original_vocal == translations["notfound"]: original_vocal = get_audio_file('Original_Vocals.')
-
-            if original_vocal == translations["notfound"]: 
-                gr_warning(translations["not_found_original_vocal"])
-                return return_none
-            
-            input_path = original_vocal
-        else:
-            main_vocal = get_audio_file('Main_Vocals_No_Reverb.')
-            backing_vocal = get_audio_file('Backing_Vocals_No_Reverb.')
-
-            if main_vocal == translations["notfound"]: main_vocal = get_audio_file('Main_Vocals.')
-            if not not_merge_backing and backing_vocal == translations["notfound"]: backing_vocal = get_audio_file('Backing_Vocals.')
-
-            if main_vocal == translations["notfound"]: 
-                gr_warning(translations["not_found_main_vocal"])
-                return return_none
-            
-            if not not_merge_backing and backing_vocal == translations["notfound"]: 
-                gr_warning(translations["not_found_backing_vocal"])
-                return return_none
-            
-            input_path = main_vocal
-            backing_path = backing_vocal
-
-        gr_info(translations["convert_vocal"])
-
-        convert(pitch, filter_radius, index_rate, volume_envelope, protect, hop_length, f0method, input_path, output_path, model_path, index, autotune, clean, clean_strength, format, embedder_model, resample_sr, split_audio, f0_autotune_strength, checkpointing, onnx_f0_mode, embedders_mode, formant_shifting, formant_qfrency, formant_timbre, f0_file)
-
-        gr_info(translations["convert_success"])
-
-        if convert_backing:
-            if os.path.exists(output_backing): os.remove(output_backing)
-
-            gr_info(translations["convert_backup"])
-
-            convert(pitch, filter_radius, index_rate, volume_envelope, protect, hop_length, f0method, backing_path, output_backing, model_path, index, autotune, clean, clean_strength, format, embedder_model, resample_sr, split_audio, f0_autotune_strength, checkpointing, onnx_f0_mode, embedders_mode, formant_shifting, formant_qfrency, formant_timbre, f0_file)
-
-            gr_info(translations["convert_backup_success"])
-
-        try:
-            if not not_merge_backing and not use_original:
-                backing_source = output_backing if convert_backing else backing_vocal
-
-                if os.path.exists(output_merge_backup): os.remove(output_merge_backup)
-
-                gr_info(translations["merge_backup"])
-
-                pydub_convert(pydub_load(output_path)).overlay(pydub_convert(pydub_load(backing_source))).export(output_merge_backup, format=format)
-
-                gr_info(translations["merge_success"])
-
-            if merge_instrument:    
-                vocals = output_merge_backup if not not_merge_backing and not use_original else output_path
-
-                if os.path.exists(output_merge_instrument): os.remove(output_merge_instrument)
-
-                gr_info(translations["merge_instruments_process"])
-
-                instruments = get_audio_file('Instruments.')
-                
-                if instruments == translations["notfound"]: 
-                    gr_warning(translations["not_found_instruments"])
-                    output_merge_instrument = None
-                else: pydub_convert(pydub_load(instruments)).overlay(pydub_convert(pydub_load(vocals))).export(output_merge_instrument, format=format)
-                
-                gr_info(translations["merge_success"])
-        except:
-            return return_none
-
-        return [(None if use_original else output_path), output_backing, (None if not_merge_backing and use_original else output_merge_backup), (output_path if use_original else None), (output_merge_instrument if merge_instrument else None), {"visible": True, "__type__": "update"}]
-    else:
-        if not input or not os.path.exists(input) or os.path.isdir(input): 
-            gr_warning(translations["input_not_valid"])
-            return return_none
-        
-        if not output:
-            gr_warning(translations["output_not_valid"])
-            return return_none
-        
-        output = output.replace("wav", format)
-
-        if os.path.isdir(input):
-            gr_info(translations["is_folder"])
-
-            if not [f for f in os.listdir(input) if f.lower().endswith(("wav", "mp3", "flac", "ogg", "opus", "m4a", "mp4", "aac", "alac", "wma", "aiff", "webm", "ac3"))]:
-                gr_warning(translations["not_found_in_folder"])
-                return return_none
-            
-            gr_info(translations["batch_convert"])
-
-            output_dir = os.path.dirname(output) or output
-            convert(pitch, filter_radius, index_rate, volume_envelope, protect, hop_length, f0method, input, output_dir, model_path, index, autotune, clean, clean_strength, format, embedder_model, resample_sr, split_audio, f0_autotune_strength, checkpointing, onnx_f0_mode, embedders_mode, formant_shifting, formant_qfrency, formant_timbre, f0_file)
-
-            gr_info(translations["batch_convert_success"])
-
-            return return_none
-        else:
-            output_dir = os.path.dirname(output) or output
-
-            if not os.path.exists(output_dir): os.makedirs(output_dir, exist_ok=True)
-            if os.path.exists(output): os.remove(output)
-
-            gr_info(translations["convert_vocal"])
-
-            convert(pitch, filter_radius, index_rate, volume_envelope, protect, hop_length, f0method, input, output, model_path, index, autotune, clean, clean_strength, format, embedder_model, resample_sr, split_audio, f0_autotune_strength, checkpointing, onnx_f0_mode, embedders_mode, formant_shifting, formant_qfrency, formant_timbre, f0_file)
-
-            gr_info(translations["convert_success"])
-
-            return_none[0] = output
-            return return_none
-
-def convert_selection(clean, autotune, use_audio, use_original, convert_backing, not_merge_backing, merge_instrument, pitch, clean_strength, model, index, index_rate, input, output, format, method, hybrid_method, hop_length, embedders, custom_embedders, resample_sr, filter_radius, volume_envelope, protect, split_audio, f0_autotune_strength, checkpointing, onnx_f0_mode, formant_shifting, formant_qfrency, formant_timbre, f0_file, embedders_mode):
-    if use_audio:
-        gr_info(translations["search_separate"])
-
-        choice = [f for f in os.listdir("audios") if os.path.isdir(os.path.join("audios", f))]
-
-        gr_info(translations["found_choice"].format(choice=len(choice)))
-
-        if len(choice) == 0: 
-            gr_warning(translations["separator==0"])
-
-            return [{"choices": [], "value": "", "interactive": False, "visible": False, "__type__": "update"}, None, None, None, None, None, {"visible": True, "__type__": "update"}]
-        elif len(choice) == 1:
-            convert_output = convert_audio(clean, autotune, use_audio, use_original, convert_backing, not_merge_backing, merge_instrument, pitch, clean_strength, model, index, index_rate, None, None, format, method, hybrid_method, hop_length, embedders, custom_embedders, resample_sr, filter_radius, volume_envelope, protect, split_audio, f0_autotune_strength, choice[0], checkpointing, onnx_f0_mode, formant_shifting, formant_qfrency, formant_timbre, f0_file, embedders_mode)
-
-            return [{"choices": [], "value": "", "interactive": False, "visible": False, "__type__": "update"}, convert_output[0], convert_output[1], convert_output[2], convert_output[3], convert_output[4], {"visible": True, "__type__": "update"}]
-        else: return [{"choices": choice, "value": "", "interactive": True, "visible": True, "__type__": "update"}, None, None, None, None, None, {"visible": False, "__type__": "update"}]
-    else:
-        main_convert = convert_audio(clean, autotune, use_audio, use_original, convert_backing, not_merge_backing, merge_instrument, pitch, clean_strength, model, index, index_rate, input, output, format, method, hybrid_method, hop_length, embedders, custom_embedders, resample_sr, filter_radius, volume_envelope, protect, split_audio, f0_autotune_strength, None, checkpointing, onnx_f0_mode, formant_shifting, formant_qfrency, formant_timbre, f0_file, embedders_mode)
-
-        return [{"choices": [], "value": "", "interactive": False, "visible": False, "__type__": "update"}, main_convert[0], None, None, None, None, {"visible": True, "__type__": "update"}]
-    
-def convert_with_whisper(num_spk, model_size, cleaner, clean_strength, autotune, f0_autotune_strength, checkpointing, model_1, model_2, model_index_1, model_index_2, pitch_1, pitch_2, index_strength_1, index_strength_2, export_format, input_audio, output_audio, onnx_f0_mode, method, hybrid_method, hop_length, embed_mode, embedders, custom_embedders, resample_sr, filter_radius, volume_envelope, protect, formant_shifting, formant_qfrency_1, formant_timbre_1, formant_qfrency_2, formant_timbre_2):
-    from pydub import AudioSegment
-    from sklearn.cluster import AgglomerativeClustering
-    
-    from main.library.speaker_diarization.audio import Audio
-    from main.library.speaker_diarization.segment import Segment
-    from main.library.speaker_diarization.whisper import load_model
-    from main.library.utils import check_spk_diarization, pydub_convert, pydub_load
-    from main.library.speaker_diarization.embedding import SpeechBrainPretrainedSpeakerEmbedding
-    
-    check_spk_diarization(model_size)
-    model_pth_1, model_pth_2 = os.path.join("assets", "weights", model_1), os.path.join("assets", "weights", model_2)
-
-    if (not model_1 or not os.path.exists(model_pth_1) or os.path.isdir(model_pth_1) or not model_pth_1.endswith((".pth", ".onnx"))) and (not model_2 or not os.path.exists(model_pth_2) or os.path.isdir(model_pth_2) or not model_pth_2.endswith((".pth", ".onnx"))):
-        gr_warning(translations["provide_file"].format(filename=translations["model"]))
-        return None
-    
-    if not model_1: model_pth_1 = model_pth_2
-    if not model_2: model_pth_2 = model_pth_1
-
-    if not input_audio or not os.path.exists(input_audio) or os.path.isdir(input_audio): 
-        gr_warning(translations["input_not_valid"])
-        return None
-        
-    if not output_audio:
-        gr_warning(translations["output_not_valid"])
-        return None
-    
-    if os.path.exists(output_audio): os.remove(output_audio)
-    gr_info(translations["start_whisper"])
-    
-    try:
-        audio = Audio()
-
-        embedding_model = SpeechBrainPretrainedSpeakerEmbedding(device=config.device)
-        segments = load_model(model_size, device=config.device).transcribe(input_audio, fp16=configs.get("fp16", False), word_timestamps=True)["segments"]
-
-        y, sr = librosa.load(input_audio, sr=None)  
-        duration = len(y) / sr
-            
-        def segment_embedding(segment):
-            waveform, _ = audio.crop(input_audio, Segment(segment["start"], min(duration, segment["end"])))
-            return embedding_model(waveform.mean(dim=0, keepdim=True)[None] if waveform.shape[0] == 2 else waveform[None])  
-        
-        def time(secs):
-            return datetime.timedelta(seconds=round(secs))
-        
-        def merge_audio(files_list, time_stamps, original_file_path, output_path, format):
-            def extract_number(filename):
-                match = re.search(r'_(\d+)', filename)
-                return int(match.group(1)) if match else 0
-
-            total_duration = len(pydub_load(original_file_path))
-            combined = AudioSegment.empty() 
-            current_position = 0 
-
-            for file, (start_i, end_i) in zip(sorted(files_list, key=extract_number), time_stamps):
-                if start_i > current_position: combined += AudioSegment.silent(duration=start_i - current_position)  
-                
-                combined += pydub_load(file)  
-                current_position = end_i
-
-            if current_position < total_duration: combined += AudioSegment.silent(duration=total_duration - current_position)
-            combined.export(output_path, format=format)
-
-            return output_path
-
-        embeddings = np.zeros(shape=(len(segments), 192))
-        for i, segment in enumerate(segments):
-            embeddings[i] = segment_embedding(segment)
-
-        labels = AgglomerativeClustering(num_spk).fit(np.nan_to_num(embeddings)).labels_
-        for i in range(len(segments)):
-            segments[i]["speaker"] = 'SPEAKER ' + str(labels[i] + 1)
-
-        merged_segments, current_text = [], []
-        current_speaker, current_start = None, None
-
-        for i, segment in enumerate(segments):
-            speaker = segment["speaker"]
-            start_time = segment["start"]
-            text = segment["text"][1:]  
-
-            if speaker == current_speaker:
-                current_text.append(text)
-                end_time = segment["end"]
-            else:
-                if current_speaker is not None: merged_segments.append({"speaker": current_speaker, "start": current_start, "end": end_time, "text": " ".join(current_text)})
-                
-                current_speaker = speaker
-                current_start = start_time
-                current_text = [text]
-                end_time = segment["end"]
-
-        if current_speaker is not None: merged_segments.append({"speaker": current_speaker, "start": current_start, "end": end_time, "text": " ".join(current_text)})
-
-        gr_info(translations["whisper_done"])
-
-        x = ""
-        for segment in merged_segments:
-            x += f"\n{segment['speaker']} {str(time(segment['start']))} - {str(time(segment['end']))}\n"
-            x += segment["text"] + "\n"
-
-        logger.info(x)
-
-        gr_info(translations["process_audio"])
-
-        audio = pydub_convert(pydub_load(input_audio))
-        output_folder = "audios_temp"
-
-        if os.path.exists(output_folder): shutil.rmtree(output_folder, ignore_errors=True)
-        for f in [output_folder, os.path.join(output_folder, "1"), os.path.join(output_folder, "2")]:
-            os.makedirs(f, exist_ok=True)
-
-        time_stamps, processed_segments = [], []
-        for i, segment in enumerate(merged_segments):
-            start_ms = int(segment["start"] * 1000) 
-            end_ms = int(segment["end"] * 1000)
-
-            index = i + 1
-
-            segment_filename = os.path.join(output_folder, "1" if i % 2 == 1 else "2", f"segment_{index}.wav")
-            audio[start_ms:end_ms].export(segment_filename, format="wav")
-
-            processed_segments.append(os.path.join(output_folder, "1" if i % 2 == 1 else "2", f"segment_{index}_output.wav"))
-            time_stamps.append((start_ms, end_ms))
-
-        f0method, embedder_model = (method if method != "hybrid" else hybrid_method), (embedders if embedders != "custom" else custom_embedders)
-
-        gr_info(translations["process_done_start_convert"])
-
-        convert(pitch_1, filter_radius, index_strength_1, volume_envelope, protect, hop_length, f0method, os.path.join(output_folder, "1"), output_folder, model_pth_1, model_index_1, autotune, cleaner, clean_strength, "wav", embedder_model, resample_sr, False, f0_autotune_strength, checkpointing, onnx_f0_mode, embed_mode, formant_shifting, formant_qfrency_1, formant_timbre_1, "")
-        convert(pitch_2, filter_radius, index_strength_2, volume_envelope, protect, hop_length, f0method, os.path.join(output_folder, "2"), output_folder, model_pth_2, model_index_2, autotune, cleaner, clean_strength, "wav", embedder_model, resample_sr, False, f0_autotune_strength, checkpointing, onnx_f0_mode, embed_mode, formant_shifting, formant_qfrency_2, formant_timbre_2, "")
-
-        gr_info(translations["convert_success"])
-        return merge_audio(processed_segments, time_stamps, input_audio, output_audio.replace("wav", export_format), export_format)
-    except Exception as e:
-        gr_error(translations["error_occurred"].format(e=e))
-        import traceback
-        logger.debug(traceback.format_exc())
-        return None
-    finally:
-        if os.path.exists("audios_temp"): shutil.rmtree("audios_temp", ignore_errors=True)
-
-def convert_tts(clean, autotune, pitch, clean_strength, model, index, index_rate, input, output, format, method, hybrid_method, hop_length, embedders, custom_embedders, resample_sr, filter_radius, volume_envelope, protect, split_audio, f0_autotune_strength, checkpointing, onnx_f0_mode, formant_shifting, formant_qfrency, formant_timbre, f0_file, embedders_mode):
-    model_path = os.path.join("assets", "weights", model)
-
-    if not model_path or not os.path.exists(model_path) or os.path.isdir(model_path) or not model.endswith((".pth", ".onnx")):
-        gr_warning(translations["provide_file"].format(filename=translations["model"]))
-        return None
-
-    if not input or not os.path.exists(input): 
-        gr_warning(translations["input_not_valid"])
-        return None
-    
-    if os.path.isdir(input): 
-        input_audio = [f for f in os.listdir(input) if "tts" in f and f.lower().endswith(("wav", "mp3", "flac", "ogg", "opus", "m4a", "mp4", "aac", "alac", "wma", "aiff", "webm", "ac3"))]
-        
-        if not input_audio:
-            gr_warning(translations["not_found_in_folder"])
-            return None
-        
-        input = os.path.join(input, input_audio[0])
-    
-    if not output:
-        gr_warning(translations["output_not_valid"])
-        return None
-    
-    output = output.replace("wav", format)
-    if os.path.isdir(output): output = os.path.join(output, f"tts.{format}")
-
-    output_dir = os.path.dirname(output)
-    if not os.path.exists(output_dir): os.makedirs(output_dir, exist_ok=True)
-    
-    if os.path.exists(output): os.remove(output)
-
-    f0method = method if method != "hybrid" else hybrid_method
-    embedder_model = embedders if embedders != "custom" else custom_embedders
-
-    gr_info(translations["convert_vocal"])
-
-    convert(pitch, filter_radius, index_rate, volume_envelope, protect, hop_length, f0method, input, output, model_path, index, autotune, clean, clean_strength, format, embedder_model, resample_sr, split_audio, f0_autotune_strength, checkpointing, onnx_f0_mode, embedders_mode, formant_shifting, formant_qfrency, formant_timbre, f0_file)
-
-    gr_info(translations["convert_success"])
-    return output
-
-def log_read(log_file, done):
-    f = open(log_file, "w", encoding="utf-8")
-    f.close()
-
-    while 1:
-        with open(log_file, "r", encoding="utf-8") as f:
-            yield "".join(line for line in f.readlines() if "DEBUG" not in line and line.strip() != "")
-
-        sleep(1)
-        if done[0]: break
-
-    with open(log_file, "r", encoding="utf-8") as f:
-        log = "".join(line for line in f.readlines() if "DEBUG" not in line and line.strip() != "")
-
-    yield log
-
-def create_dataset(input_audio, output_dataset, clean_dataset, clean_strength, separator_reverb, kim_vocals_version, overlap, segments_size, denoise_mdx, skip, skip_start, skip_end, hop_length, batch_size, sample_rate):
-    version = 1 if kim_vocals_version == "Version-1" else 2
-
-    gr_info(translations["start"].format(start=translations["create"]))
-
-    p = subprocess.Popen(f'{python} main/inference/create_dataset.py --input_audio "{input_audio}" --output_dataset "{output_dataset}" --clean_dataset {clean_dataset} --clean_strength {clean_strength} --separator_reverb {separator_reverb} --kim_vocal_version {version} --overlap {overlap} --segments_size {segments_size} --mdx_hop_length {hop_length} --mdx_batch_size {batch_size} --denoise_mdx {denoise_mdx} --skip {skip} --skip_start_audios "{skip_start}" --skip_end_audios "{skip_end}" --sample_rate {sample_rate}', shell=True)
-    done = [False]
-
-    threading.Thread(target=if_done, args=(done, p)).start()
-
-    for log in log_read(os.path.join("assets", "logs", "create_dataset.log"), done):
-        yield log
-
-def preprocess(model_name, sample_rate, cpu_core, cut_preprocess, process_effects, path, clean_dataset, clean_strength):
-    dataset = os.path.join(path)
-    sr = int(float(sample_rate.rstrip("k")) * 1000)
-
-    if not model_name: return gr_warning(translations["provide_name"])
-    if not any(f.lower().endswith(("wav", "mp3", "flac", "ogg", "opus", "m4a", "mp4", "aac", "alac", "wma", "aiff", "webm", "ac3")) for f in os.listdir(dataset) if os.path.isfile(os.path.join(dataset, f))): return gr_warning(translations["not_found_data"])
-    
-    model_dir = os.path.join("assets", "logs", model_name)
-    if os.path.exists(model_dir): shutil.rmtree(model_dir, ignore_errors=True)
-
-    p = subprocess.Popen(f'{python} main/inference/preprocess.py --model_name "{model_name}" --dataset_path "{dataset}" --sample_rate {sr} --cpu_cores {cpu_core} --cut_preprocess {cut_preprocess} --process_effects {process_effects} --clean_dataset {clean_dataset} --clean_strength {clean_strength}', shell=True)
-    done = [False]
-
-    threading.Thread(target=if_done, args=(done, p)).start()
-    os.makedirs(model_dir, exist_ok=True)
-
-    for log in log_read(os.path.join(model_dir, "preprocess.log"), done):
-        yield log
-
-def extract(model_name, version, method, pitch_guidance, hop_length, cpu_cores, gpu, sample_rate, embedders, custom_embedders, onnx_f0_mode, embedders_mode):
-    embedder_model = embedders if embedders != "custom" else custom_embedders
-    sr = int(float(sample_rate.rstrip("k")) * 1000)
-
-    if not model_name: return gr_warning(translations["provide_name"])
-
-    model_dir = os.path.join("assets", "logs", model_name)
-    if not any(os.path.isfile(os.path.join(model_dir, "sliced_audios", f)) for f in os.listdir(os.path.join(model_dir, "sliced_audios"))) or not any(os.path.isfile(os.path.join(model_dir, "sliced_audios_16k", f)) for f in os.listdir(os.path.join(model_dir, "sliced_audios_16k"))): return gr_warning(translations["not_found_data_preprocess"])
-
-    p = subprocess.Popen(f'{python} main/inference/extract.py --model_name "{model_name}" --rvc_version {version} --f0_method {method} --pitch_guidance {pitch_guidance} --hop_length {hop_length} --cpu_cores {cpu_cores} --gpu {gpu} --sample_rate {sr} --embedder_model {embedder_model} --f0_onnx {onnx_f0_mode} --embedders_mode {embedders_mode}', shell=True)
-    done = [False]
-
-    threading.Thread(target=if_done, args=(done, p)).start()
-    os.makedirs(model_dir, exist_ok=True)
-
-    for log in log_read(os.path.join(model_dir, "extract.log"), done):
-        yield log
-
-def create_index(model_name, rvc_version, index_algorithm):
-    if not model_name: return gr_warning(translations["provide_name"])
-    model_dir = os.path.join("assets", "logs", model_name)
-
-    if not any(os.path.isfile(os.path.join(model_dir, f"{rvc_version}_extracted", f)) for f in os.listdir(os.path.join(model_dir, f"{rvc_version}_extracted"))): return gr_warning(translations["not_found_data_extract"])
-
-    p = subprocess.Popen(f'{python} main/inference/create_index.py --model_name "{model_name}" --rvc_version {rvc_version} --index_algorithm {index_algorithm}', shell=True)
-    done = [False]
-
-    threading.Thread(target=if_done, args=(done, p)).start()
-    os.makedirs(model_dir, exist_ok=True)
-
-    for log in log_read(os.path.join(model_dir, "create_index.log"), done):
-        yield log
-
-def training(model_name, rvc_version, save_every_epoch, save_only_latest, save_every_weights, total_epoch, sample_rate, batch_size, gpu, pitch_guidance, not_pretrain, custom_pretrained, pretrain_g, pretrain_d, detector, threshold, clean_up, cache, model_author, vocoder, checkpointing, deterministic, benchmark):
-    sr = int(float(sample_rate.rstrip("k")) * 1000)
-    if not model_name: return gr_warning(translations["provide_name"])
-
-    model_dir = os.path.join("assets", "logs", model_name)
-    if os.path.exists(os.path.join(model_dir, "train_pid.txt")): os.remove(os.path.join(model_dir, "train_pid.txt"))
-
-    if not any(os.path.isfile(os.path.join(model_dir, f"{rvc_version}_extracted", f)) for f in os.listdir(os.path.join(model_dir, f"{rvc_version}_extracted"))): return gr_warning(translations["not_found_data_extract"])
-
-    if not not_pretrain:
-        if not custom_pretrained: 
-            pretrained_selector = {True: {32000: ("f0G32k.pth", "f0D32k.pth"), 40000: ("f0G40k.pth", "f0D40k.pth"), 48000: ("f0G48k.pth", "f0D48k.pth")}, False: {32000: ("G32k.pth", "D32k.pth"), 40000: ("G40k.pth", "D40k.pth"), 48000: ("G48k.pth", "D48k.pth")}}
-
-            pg, pd = pretrained_selector[pitch_guidance][sr]
-        else:
-            if not pretrain_g: return gr_warning(translations["provide_pretrained"].format(dg="G"))
-            if not pretrain_d: return gr_warning(translations["provide_pretrained"].format(dg="D"))
-            
-            pg, pd = pretrain_g, pretrain_d
-
-        pretrained_G, pretrained_D = (os.path.join("assets", "models", f"pretrained_{rvc_version}", f"{vocoder}_{pg}" if vocoder != 'Default' else pg), os.path.join("assets", "models", f"pretrained_{rvc_version}", f"{vocoder}_{pd}" if vocoder != 'Default' else pd)) if not custom_pretrained else (os.path.join("assets", "models", f"pretrained_custom", pg), os.path.join("assets", "models", f"pretrained_custom", pd))
-        download_version = codecs.decode(f"uggcf://uhttvatsnpr.pb/NauC/Ivrganzrfr-EIP-Cebwrpg/erfbyir/znva/cergenvarq_i{'2' if rvc_version == 'v2' else '1'}/", "rot13")
-        
-        if not custom_pretrained:
-            try:
-                if not os.path.exists(pretrained_G):
-                    gr_info(translations["download_pretrained"].format(dg="G", rvc_version=rvc_version))
-                    huggingface.HF_download_file("".join([download_version, vocoder, "_", pg]) if vocoder != 'Default' else (download_version + pg), os.path.join("assets", "models", f"pretrained_{rvc_version}", f"{vocoder}_{pg}" if vocoder != 'Default' else pg))
-                        
-                if not os.path.exists(pretrained_D):
-                    gr_info(translations["download_pretrained"].format(dg="D", rvc_version=rvc_version))
-                    huggingface.HF_download_file("".join([download_version, vocoder, "_", pd]) if vocoder != 'Default' else (download_version + pd), os.path.join("assets", "models", f"pretrained_{rvc_version}", f"{vocoder}_{pd}" if vocoder != 'Default' else pd))
-            except:
-                gr_warning(translations["not_use_pretrain_error_download"])
-                pretrained_G, pretrained_D = None, None
-        else:
-            if not os.path.exists(pretrained_G): return gr_warning(translations["not_found_pretrain"].format(dg="G"))
-            if not os.path.exists(pretrained_D): return gr_warning(translations["not_found_pretrain"].format(dg="D"))
-    else: gr_warning(translations["not_use_pretrain"])
-
-    gr_info(translations["start"].format(start=translations["training"]))
-
-    p = subprocess.Popen(f'{python} main/inference/train.py --model_name "{model_name}" --rvc_version {rvc_version} --save_every_epoch {save_every_epoch} --save_only_latest {save_only_latest} --save_every_weights {save_every_weights} --total_epoch {total_epoch} --sample_rate {sr} --batch_size {batch_size} --gpu {gpu} --pitch_guidance {pitch_guidance} --overtraining_detector {detector} --overtraining_threshold {threshold} --cleanup {clean_up} --cache_data_in_gpu {cache} --g_pretrained_path "{pretrained_G}" --d_pretrained_path "{pretrained_D}" --model_author "{model_author}" --vocoder "{vocoder}" --checkpointing {checkpointing} --deterministic {deterministic} --benchmark {benchmark}', shell=True)
-    done = [False]
-
-    with open(os.path.join(model_dir, "train_pid.txt"), "w") as pid_file:
-        pid_file.write(str(p.pid))
-
-    threading.Thread(target=if_done, args=(done, p)).start()
-
-    for log in log_read(os.path.join(model_dir, "train.log"), done):
-        if len(log.split("\n")) > 100: log = log[-100:]
-        yield log
-
-def stop_pid(pid_file, model_name=None, train=False):
-    try:
-        pid_file_path = os.path.join("assets", f"{pid_file}.txt") if model_name is None else os.path.join("assets", "logs", model_name, f"{pid_file}.txt")
-
-        if not os.path.exists(pid_file_path): return gr_warning(translations["not_found_pid"])
-        else:
-            with open(pid_file_path, "r") as pid_file:
-                pids = [int(pid) for pid in pid_file.readlines()]
-
-            for pid in pids:
-                os.kill(pid, 9)
-
-            if os.path.exists(pid_file_path): os.remove(pid_file_path)
-
-        pid_file_path = os.path.join("assets", "logs", model_name, "config.json")
-
-        if train and os.path.exists(pid_file_path):
-            with open(pid_file_path, "r") as pid_file:
-                pid_data = json.load(pid_file)
-                pids = pid_data.get("process_pids", [])
-
-            with open(pid_file_path, "w") as pid_file:
-                pid_data.pop("process_pids", None)
-
-                json.dump(pid_data, pid_file, indent=4)
-
-            for pid in pids:
-                os.kill(pid, 9)
-
-            gr_info(translations["end_pid"])
-    except:
-        pass
-
-def load_presets(presets, cleaner, autotune, pitch, clean_strength, index_strength, resample_sr, filter_radius, volume_envelope, protect, split_audio, f0_autotune_strength, formant_shifting, formant_qfrency, formant_timbre):
-    if not presets: return gr_warning(translations["provide_file_settings"])
-
-    with open(os.path.join("assets", "presets", presets)) as f:
-        file = json.load(f)
-
-    gr_info(translations["load_presets"].format(presets=presets))
-    return file.get("cleaner", cleaner), file.get("autotune", autotune), file.get("pitch", pitch), file.get("clean_strength", clean_strength), file.get("index_strength", index_strength), file.get("resample_sr", resample_sr), file.get("filter_radius", filter_radius), file.get("volume_envelope", volume_envelope), file.get("protect", protect), file.get("split_audio", split_audio), file.get("f0_autotune_strength", f0_autotune_strength), file.get("formant_shifting", formant_shifting), file.get("formant_qfrency", formant_qfrency), file.get("formant_timbre", formant_timbre)
-
-def save_presets(name, cleaner, autotune, pitch, clean_strength, index_strength, resample_sr, filter_radius, volume_envelope, protect, split_audio, f0_autotune_strength, cleaner_chbox, autotune_chbox, pitch_chbox, index_strength_chbox, resample_sr_chbox, filter_radius_chbox, volume_envelope_chbox, protect_chbox, split_audio_chbox, formant_shifting_chbox, formant_shifting, formant_qfrency, formant_timbre):  
-    if not name: return gr_warning(translations["provide_filename_settings"])
-    if not any([cleaner_chbox, autotune_chbox, pitch_chbox, index_strength_chbox, resample_sr_chbox, filter_radius_chbox, volume_envelope_chbox, protect_chbox, split_audio_chbox, formant_shifting_chbox]): return gr_warning(translations["choose1"])
-
-    settings = {}
-
-    for checkbox, data in [(cleaner_chbox, {"cleaner": cleaner, "clean_strength": clean_strength}), (autotune_chbox, {"autotune": autotune, "f0_autotune_strength": f0_autotune_strength}), (pitch_chbox, {"pitch": pitch}), (index_strength_chbox, {"index_strength": index_strength}), (resample_sr_chbox, {"resample_sr": resample_sr}), (filter_radius_chbox, {"filter_radius": filter_radius}), (volume_envelope_chbox, {"volume_envelope": volume_envelope}), (protect_chbox, {"protect": protect}), (split_audio_chbox, {"split_audio": split_audio}), (formant_shifting_chbox, {"formant_shifting": formant_shifting, "formant_qfrency": formant_qfrency, "formant_timbre": formant_timbre})]:
-        if checkbox: settings.update(data)
-
-    with open(os.path.join("assets", "presets", name + ".json"), "w") as f:
-        json.dump(settings, f, indent=4)
-
-    gr_info(translations["export_settings"])
-    return change_preset_choices()
-
-def report_bug(error_info, provide):
-    report_path = os.path.join("assets", "logs", "report_bugs.log")
-    if os.path.exists(report_path): os.remove(report_path)
-
-    report_url = codecs.decode(requests.get(codecs.decode("uggcf://uhttvatsnpr.pb/NauC/Ivrganzrfr-EIP-Cebwrpg/erfbyir/znva/jroubbx.gkg", "rot13")).text, "rot13")
-    if not error_info: error_info = "Không Có"
-
-    gr_info(translations["thank"])
-
-    if provide:
-        try:
-            for log in [os.path.join(root, name) for root, _, files in os.walk(os.path.join("assets", "logs"), topdown=False) for name in files if name.endswith(".log")]:
-                with open(log, "r", encoding="utf-8") as r:
-                    with open(report_path, "a", encoding="utf-8") as w:
-                        w.write(str(r.read()))
-                        w.write("\n")
-        except Exception as e:
-            gr_error(translations["error_read_log"])
-            logger.debug(e)
-
-        try:
-            with open(report_path, "r", encoding="utf-8") as f:
-                content = f.read()
-
-            requests.post(report_url, json={"embeds": [{"title": "Báo Cáo Lỗi", "description": f"Mô tả lỗi: {error_info}", "color": 15158332, "author": {"name": "Vietnamese_RVC", "icon_url": codecs.decode("uggcf://uhttvatsnpr.pb/NauC/Ivrganzrfr-EIP-Cebwrpg/erfbyir/znva/vpb.cat", "rot13"), "url": codecs.decode("uggcf://tvguho.pbz/CunzUhlauNau16/Ivrganzrfr-EIP/gerr/znva","rot13")}, "thumbnail": {"url": codecs.decode("uggcf://p.grabe.pbz/7dADJbv-36fNNNNq/grabe.tvs", "rot13")}, "fields": [{"name": "Số Lượng Gỡ Lỗi", "value": content.count("DEBUG")}, {"name": "Số Lượng Thông Tin", "value": content.count("INFO")}, {"name": "Số Lượng Cảnh Báo", "value": content.count("WARNING")}, {"name": "Số Lượng Lỗi", "value": content.count("ERROR")}], "footer": {"text": f"Tên Máy: {platform.uname().node} - Hệ Điều Hành: {platform.system()}-{platform.version()}\nThời Gian Báo Cáo Lỗi: {datetime.datetime.now()}."}}]})
-
-            with open(report_path, "rb") as f:
-                requests.post(report_url, files={"file": f})
-        except Exception as e:
-            gr_error(translations["error_send"])
-            logger.debug(e)
-        finally:
-            if os.path.exists(report_path): os.remove(report_path)
-    else: requests.post(report_url, json={"embeds": [{"title": "Báo Cáo Lỗi", "description": error_info}]})
-
-def f0_extract(audio, f0_method, f0_onnx):
-    if not audio or not os.path.exists(audio) or os.path.isdir(audio): 
-        gr_warning(translations["input_not_valid"])
-        return [None]*2
-
-    from matplotlib import pyplot as plt
-    from main.library.utils import check_predictors
-    from main.inference.extract import FeatureInput
-
-    check_predictors(f0_method, f0_onnx)
-
-    f0_path = os.path.join("assets", "f0", os.path.splitext(os.path.basename(audio))[0])
-    image_path = os.path.join(f0_path, "f0.png")
-    txt_path = os.path.join(f0_path, "f0.txt")
-
-    gr_info(translations["start_extract"])
-
-    if not os.path.exists(f0_path): os.makedirs(f0_path, exist_ok=True)
-
-    y, sr = librosa.load(audio, sr=None)
-
-    feats = FeatureInput(sample_rate=sr, is_half=config.is_half, device=config.device)
-    feats.f0_max = 1600.0
-
-    F_temp = np.array(feats.compute_f0(y.flatten(), f0_method, 160, f0_onnx), dtype=np.float32)
-    F_temp[F_temp == 0] = np.nan
-
-    f0 = 1200 * np.log2(F_temp / librosa.midi_to_hz(0))
-
-    plt.figure(figsize=(10, 4))
-    plt.plot(f0)
-    plt.title(f0_method)
-    plt.xlabel(translations["time_frames"])
-    plt.ylabel(translations["Frequency"])
-    plt.savefig(image_path)
-    plt.close()
-
-    with open(txt_path, "w") as f:
-        for i, f0_value in enumerate(f0):
-            f.write(f"{i * sr / 160},{f0_value}\n")
-
-    gr_info(translations["extract_done"])
-
-    return [txt_path, image_path]
-
-def pitch_guidance_lock(vocoders):
-    return {"value": True, "interactive": vocoders == "Default", "__type__": "update"}
-
-def vocoders_lock(pitch, vocoders):
-    return {"value": vocoders if pitch else "Default", "interactive": pitch, "__type__": "update"}
-
-def run_audioldm2(input_path, output_path, export_format, sample_rate, audioldm_model, source_prompt, target_prompt, steps, cfg_scale_src, cfg_scale_tar, t_start, save_compute):
-    if not input_path or not os.path.exists(input_path) or os.path.isdir(input_path): 
-        gr_warning(translations["input_not_valid"])
-        return None
-        
-    if not output_path:
-        gr_warning(translations["output_not_valid"])
-        return None
-    
-    output_path = output_path.replace("wav", export_format)
-
-    if os.path.exists(output_path): os.remove(output_path)
-
-    gr_info(translations["start_edit"].format(input_path=input_path))
-    subprocess.run([python, "main/inference/audioldm2.py", "--input_path", input_path, "--output_path", output_path, "--export_format", str(export_format), "--sample_rate", str(sample_rate), "--audioldm_model", audioldm_model, "--source_prompt", source_prompt, "--target_prompt", target_prompt, "--steps", str(steps), "--cfg_scale_src", str(cfg_scale_src), "--cfg_scale_tar", str(cfg_scale_tar), "--t_start", str(t_start), "--save_compute", str(save_compute)])
-    
-    gr_info(translations["success"])
-    return output_path
-
-def change_fp(fp):
-    fp16 = fp == "fp16"
-
-    if fp16 and config.device == "cpu": 
-        gr_warning(translations["fp16_not_support"])
-        return "fp32"
-    else:
-        gr_info(translations["start_update_precision"])
-
-        configs = json.load(open(configs_json, "r"))
-        configs["fp16"] = config.is_half = fp16
-
-        with open(configs_json, "w") as f:
-            json.dump(configs, f, indent=4)
-
-        gr_info(translations["success"])
-        return "fp16" if fp16 else "fp32"
-
-def unlock_f0(value):
-    return {"choices": method_f0_full if value else method_f0, "value": "rmvpe", "__type__": "update"} 
-
-def unlock_vocoder(value, vocoder):
-    return {"value": vocoder if value == "v2" else "Default", "interactive": value == "v2", "__type__": "update"} 
-
-def unlock_ver(value, vocoder):
-    return {"value": "v2" if vocoder == "Default" else value, "interactive": vocoder == "Default", "__type__": "update"}
-
-def visible_embedders(value):
-    return {"visible": value != "spin", "__type__": "update"}
-

main/app/parser.py

@@ -1,339 +0,0 @@
-import os
-import sys
-
-sys.path.append(os.getcwd())
-
-try:
-    argv = sys.argv[1]
-except IndexError:
-    argv = None
-
-argv_is_allows = ["--audio_effects", "--audioldm2", "--convert", "--create_dataset", "--create_index", "--extract", "--preprocess", "--separator_music", "--train", "--help_audio_effects", "--help_audioldm2", "--help_convert", "--help_create_dataset", "--help_create_index", "--help_extract", "--help_preprocess", "--help_separator_music", "--help_train", "--help"]
-
-if argv not in argv_is_allows:
-    print("Invalid syntax! Use --help for more information")
-    quit()
-
-if argv_is_allows[0] in argv: from main.inference.audio_effects import main
-elif argv_is_allows[1] in argv: from main.inference.audioldm2 import main
-elif argv_is_allows[2] in argv: from main.inference.convert import main
-elif argv_is_allows[3] in argv: from main.inference.create_dataset import main
-elif argv_is_allows[4] in argv: from main.inference.create_index import main
-elif argv_is_allows[5] in argv: from main.inference.extract import main
-elif argv_is_allows[6] in argv: from main.inference.preprocess import main
-elif argv_is_allows[7] in argv: from main.inference.separator_music import main
-elif argv_is_allows[8] in argv: from main.inference.train import main
-elif argv_is_allows[9] in argv:
-    print("""Parameters for `--audio_effects`:
-        1. File paths:
-            - `--input_path` (required): Path to the input audio file.
-            - `--output_path` (default: `./audios/apply_effects.wav`): Path to save the output file.
-            - `--export_format` (default: `wav`): Output file format (`wav`, `mp3`, ...).
-
-        2. Resampling:
-            - `--resample` (default: `False`): Whether to resample or not.
-            - `--resample_sr` (default: `0`): New sampling frequency (Hz).
-
-        3. Chorus effect:
-            - `--chorus`: Enable/disable chorus.
-            - `--chorus_depth`, `--chorus_rate`, `--chorus_mix`, `--chorus_delay`, `--chorus_feedback`: Parameters to adjust chorus.
-
-        4. Distortion effect:
-            - `--distortion`: Enable/disable distortion.
-            - `--drive_db`: Degree of audio distortion.
-
-        5. Reverb effect:
-            - `--reverb`: Enable/disable reverb.
-            - `--reverb_room_size`, `--reverb_damping`, `--reverb_wet_level`, `--reverb_dry_level`, `--reverb_width`, `--reverb_freeze_mode`: Adjust reverb.
-
-        6. Pitch shift effect:
-            - `--pitchshift`: Enable/disable pitch shift.
-            - `--pitch_shift`: Pitch shift value.
-
-        7. Delay effect:
-            - `--delay`: Enable/disable delay.
-            - `--delay_seconds`, `--delay_feedback`, `--delay_mix`: Adjust delay time, feedback, and mix.
-
-        8. Compressor:
-            - `--compressor`: Enable/disable compressor.
-            - `--compressor_threshold`, `--compressor_ratio`, `--compressor_attack_ms`, `--compressor_release_ms`: Compression parameters.
-
-        9. Limiter:
-            - `--limiter`: Enable/disable audio level limiter.
-            - `--limiter_threshold`, `--limiter_release`: Limiter threshold and release time.
-
-        10. Gain (Amplification):
-            - `--gain`: Enable/disable gain.
-            - `--gain_db`: Gain level (dB).
-
-        11. Bitcrush:
-            - `--bitcrush`: Enable/disable bit resolution reduction effect.
-            - `--bitcrush_bit_depth`: Bit depth for bitcrush.
-
-        12. Clipping:
-            - `--clipping`: Enable/disable audio clipping.
-            - `--clipping_threshold`: Clipping threshold.
-
-        13. Phaser:
-            - `--phaser`: Enable/disable phaser effect.
-            - `--phaser_rate_hz`, `--phaser_depth`, `--phaser_centre_frequency_hz`, `--phaser_feedback`, `--phaser_mix`: Adjust phaser effect.
-
-        14. Boost bass & treble:
-            - `--treble_bass_boost`: Enable/disable bass and treble boost.
-            - `--bass_boost_db`, `--bass_boost_frequency`, `--treble_boost_db`, `--treble_boost_frequency`: Bass and treble boost parameters.
-
-        15. Fade in & fade out:
-            - `--fade_in_out`: Enable/disable fade effect.
-            - `--fade_in_duration`, `--fade_out_duration`: Fade in/out duration.
-
-        16. Audio combination:
-            - `--audio_combination`: Enable/disable combining multiple audio files.
-            - `--audio_combination_input`: Path to additional audio files.
-    """)
-    quit()
-elif argv_is_allows[10] in argv:
-    print("""Parameters for `--audioldm2`:
-        1. File paths:
-            - `--input_path` (required): Path to the input audio file.
-            - `--output_path` (default: `./output.wav`): Path to save the output file.
-            - `--export_format` (default: `wav`): Output file format.
-
-        2. Audio configuration:
-            - `--sample_rate` (default: `44100`): Sampling frequency (Hz).
-
-        3. AudioLDM model configuration:
-            - `--audioldm_model` (default: `audioldm2-music`): Select AudioLDM model for processing.
-
-        4. Model guidance prompt:
-            - `--source_prompt` (default: ``): Description of source audio.
-            - `--target_prompt` (default: ``): Description of target audio.
-
-        5. Processing algorithm configuration:
-            - `--steps` (default: `200`): Number of steps in audio synthesis process.
-            - `--cfg_scale_src` (default: `3.5`): Guidance scale for source audio.
-            - `--cfg_scale_tar` (default: `12`): Guidance scale for target audio.
-            - `--t_start` (default: `45`): Editing level.
-
-        6. Computation optimization:
-            - `--save_compute` (default: `False`): Whether to enable compute optimization mode.
-    """)
-    quit()
-elif argv_is_allows[11] in argv:
-    print("""Parameters for `--convert`:
-        1. Voice processing configuration:
-            - `--pitch` (default: `0`): Adjust pitch.
-            - `--filter_radius` (default: `3`): F0 curve smoothness.
-            - `--index_rate` (default: `0.5`): Voice index usage rate.
-            - `--volume_envelope` (default: `1`): Volume amplitude adjustment factor.
-            - `--protect` (default: `0.33`): Consonant protection.
-
-        2. Frame hop configuration:
-            - `--hop_length` (default: `64`): Hop length during audio processing.
-
-        3. F0 configuration:
-            - `--f0_method` (default: `rmvpe`): F0 prediction method (`pm`, `dio`, `mangio-crepe-tiny`, `mangio-crepe-small`, `mangio-crepe-medium`, `mangio-crepe-large`, `mangio-crepe-full`, `crepe-tiny`, `crepe-small`, `crepe-medium`, `crepe-large`, `crepe-full`, `fcpe`, `fcpe-legacy`, `rmvpe`, `rmvpe-legacy`, `harvest`, `yin`, `pyin`, `swipe`).
-            - `--f0_autotune` (default: `False`): Whether to auto-tune F0.
-            - `--f0_autotune_strength` (default: `1`): Strength of F0 auto-tuning.
-            - `--f0_file` (default: ``): Path to existing F0 file.
-            - `--f0_onnx` (default: `False`): Whether to use ONNX version of F0.
-
-        4. Embedding model:
-            - `--embedder_model` (default: `contentvec_base`): Embedding model used.
-            - `--embedders_mode` (default: `fairseq`): Embedding mode (`fairseq`, `transformers`, `onnx`).
-
-        5. File paths:
-            - `--input_path` (required): Path to input audio file.
-            - `--output_path` (default: `./audios/output.wav`): Path to save output file.
-            - `--export_format` (default: `wav`): Output file format.
-            - `--pth_path` (required): Path to `.pth` model file.
-            - `--index_path` (default: `None`): Path to index file (if any).
-
-        6. Audio cleaning:
-            - `--clean_audio` (default: `False`): Whether to apply audio cleaning.
-            - `--clean_strength` (default: `0.7`): Cleaning strength.
-
-        7. Resampling & audio splitting:
-            - `--resample_sr` (default: `0`): New sampling frequency (0 means keep original).
-            - `--split_audio` (default: `False`): Whether to split audio before processing.
-
-        8. Testing & optimization:
-            - `--checkpointing` (default: `False`): Enable/disable checkpointing to save RAM.
-
-        9. Formant shifting:
-            - `--formant_shifting` (default: `False`): Whether to enable formant shifting effect.
-            - `--formant_qfrency` (default: `0.8`): Formant shift frequency factor.
-            - `--formant_timbre` (default: `0.8`): Voice timbre change factor.
-    """)
-    quit()
-elif argv_is_allows[12] in argv:
-    print("""Parameters for `--create_dataset`:
-        1. Dataset paths & configuration:
-            - `--input_audio` (required): Path to audio link (YouTube link, can use `,` for multiple links).
-            - `--output_dataset` (default: `./dataset`): Output data directory.
-            - `--sample_rate` (default: `44100`): Audio sampling frequency.
-
-        2. Data cleaning:
-            - `--clean_dataset` (default: `False`): Whether to apply data cleaning.
-            - `--clean_strength` (default: `0.7`): Data cleaning strength.
-
-        3. Voice separation & effects:
-            - `--separator_reverb` (default: `False`): Whether to separate voice reverb.
-            - `--kim_vocal_version` (default: `2`): Kim Vocal model version for separation (`1`, `2`).
-
-        4. Audio segmentation configuration:
-            - `--overlap` (default: `0.25`): Overlap level between segments during separation.
-            - `--segments_size` (default: `256`): Size of each segment.
-
-        5. MDX (Music Demixing) configuration:
-            - `--mdx_hop_length` (default: `1024`): MDX hop length during processing.
-            - `--mdx_batch_size` (default: `1`): Batch size during MDX processing.
-            - `--denoise_mdx` (default: `False`): Whether to apply denoising during MDX separation.
-
-        6. Skip audio sections:
-            - `--skip` (default: `False`): Whether to skip any audio seconds.
-            - `--skip_start_audios` (default: `0`): Time (seconds) to skip at the start of audio.
-            - `--skip_end_audios` (default: `0`): Time (seconds) to skip at the end of audio.
-    """)
-    quit()
-elif argv_is_allows[13] in argv:
-    print("""Parameters for `--create_index`:
-        1. Model information:
-            - `--model_name` (required): Model name.
-            - `--rvc_version` (default: `v2`): Version (`v1`, `v2`).
-            - `--index_algorithm` (default: `Auto`): Index algorithm used (`Auto`, `Faiss`, `KMeans`).
-    """)
-    quit()
-elif argv_is_allows[14] in argv:
-    print("""Parameters for `--extract`:
-        1. Model information:
-            - `--model_name` (required): Model name.
-            - `--rvc_version` (default: `v2`): RVC version (`v1`, `v2`).
-
-        2. F0 configuration:
-            - `--f0_method` (default: `rmvpe`): F0 prediction method (`pm`, `dio`, `mangio-crepe-tiny`, `mangio-crepe-small`, `mangio-crepe-medium`, `mangio-crepe-large`, `mangio-crepe-full`, `crepe-tiny`, `crepe-small`, `crepe-medium`, `crepe-large`, `crepe-full`, `fcpe`, `fcpe-legacy`, `rmvpe`, `rmvpe-legacy`, `harvest`, `yin`, `pyin`, `swipe`).
-            - `--pitch_guidance` (default: `True`): Whether to use pitch guidance.
-
-        3. Processing configuration:
-            - `--hop_length` (default: `128`): Hop length during processing.
-            - `--cpu_cores` (default: `2`): Number of CPU threads used.
-            - `--gpu` (default: `-`): Specify GPU to use (e.g., `0` for first GPU, `-` to disable GPU).
-            - `--sample_rate` (required): Input audio sampling frequency.
-
-        4. Embedding configuration:
-            - `--embedder_model` (default: `contentvec_base`): Embedding model name.
-            - `--f0_onnx` (default: `False`): Whether to use ONNX version of F0.
-            - `--embedders_mode` (default: `fairseq`): Embedding mode (`fairseq`, `transformers`, `onnx`).
-    """)
-    quit()
-elif argv_is_allows[15] in argv:
-    print("""Parameters for `--preprocess`:
-        1. Model information:
-            - `--model_name` (required): Model name.
-
-        2. Data configuration:
-            - `--dataset_path` (default: `./dataset`): Path to directory containing data files.
-            - `--sample_rate` (required): Audio data sampling frequency.
-
-        3. Processing configuration:
-            - `--cpu_cores` (default: `2`): Number of CPU threads used.
-            - `--cut_preprocess` (default: `True`): Whether to cut data files.
-            - `--process_effects` (default: `False`): Whether to apply preprocessing.
-            - `--clean_dataset` (default: `False`): Whether to clean data files.
-            - `--clean_strength` (default: `0.7`): Data cleaning strength.
-    """)
-    quit()
-elif argv_is_allows[16] in argv:
-    print("""Parameters for `--separator_music`:
-        1. Data paths:
-            - `--input_path` (required): Path to input audio file.
-            - `--output_path` (default: `./audios`): Directory to save output files.
-            - `--format` (default: `wav`): Output file format (`wav`, `mp3`, ...).
-
-        2. Audio processing configuration:
-            - `--shifts` (default: `2`): Number of predictions.
-            - `--segments_size` (default: `256`): Audio segment size.
-            - `--overlap` (default: `0.25`): Overlap level between segments.
-            - `--mdx_hop_length` (default: `1024`): MDX hop length during processing.
-            - `--mdx_batch_size` (default: `1`): Batch size.
-
-        3. Cleaning processing:
-            - `--clean_audio` (default: `False`): Whether to clean audio.
-            - `--clean_strength` (default: `0.7`): Cleaning filter strength.
-
-        4. Model configuration:
-            - `--model_name` (default: `HT-Normal`): Music separation model (`Main_340`, `Main_390`, `Main_406`, `Main_427`, `Main_438`, `Inst_full_292`, `Inst_HQ_1`, `Inst_HQ_2`, `Inst_HQ_3`, `Inst_HQ_4`, `Inst_HQ_5`, `Kim_Vocal_1`, `Kim_Vocal_2`, `Kim_Inst`, `Inst_187_beta`, `Inst_82_beta`, `Inst_90_beta`, `Voc_FT`, `Crowd_HQ`, `Inst_1`, `Inst_2`, `Inst_3`, `MDXNET_1_9703`, `MDXNET_2_9682`, `MDXNET_3_9662`, `Inst_Main`, `MDXNET_Main`, `MDXNET_9482`, `HT-Normal`, `HT-Tuned`, `HD_MMI`, `HT_6S`).
-            - `--kara_model` (default: `Version-1`): Backing track separation model version (`Version-1`, `Version-2`).
-
-        5. Effects and post-processing:
-            - `--backing` (default: `False`): Whether to separate backing vocals.
-            - `--mdx_denoise` (default: `False`): Whether to use MDX denoising.
-            - `--reverb` (default: `False`): Whether to separate reverb.
-            - `--backing_reverb` (default: `False`): Whether to separate reverb for backing vocals.
-
-        6. Sampling frequency:
-            - `--sample_rate` (default: `44100`): Output audio sampling frequency.
-    """)
-    quit()
-elif argv_is_allows[17] in argv:
-    print("""Parameters for `--train`:
-        1. Model configuration:
-            - `--model_name` (required): Model name.
-            - `--rvc_version` (default: `v2`): RVC version (`v1`, `v2`).
-            - `--model_author` (optional): Model author.
-
-        2. Save configuration:
-            - `--save_every_epoch` (required): Number of epochs between saves.
-            - `--save_only_latest` (default: `True`): Save only the latest checkpoint.
-            - `--save_every_weights` (default: `True`): Save all model weights.
-
-        3. Training configuration:
-            - `--total_epoch` (default: `300`): Total number of training epochs.
-            - `--batch_size` (default: `8`): Batch size during training.
-            - `--sample_rate` (required): Audio sampling frequency.
-
-        4. Device configuration:
-            - `--gpu` (default: `0`): Specify GPU to use (agena: Specify GPU to use (number or `-` if not using GPU).
-            - `--cache_data_in_gpu` (default: `False`): Cache data in GPU for faster processing.
-
-        5. Advanced training configuration:
-            - `--pitch_guidance` (default: `True`): Use pitch guidance.
-            - `--g_pretrained_path` (default: ``): Path to pretrained G weights.
-            - `--d_pretrained_path` (default: ``): Path to pretrained D weights.
-            - `--vocoder` (default: `Default`): Vocoder used (`Default`, `MRF-HiFi-GAN`, `RefineGAN`).
-
-        6. Overtraining detection:
-            - `--overtraining_detector` (default: `False`): Enable/disable overtraining detection.
-            - `--overtraining_threshold` (default: `50`): Threshold for detecting overtraining.
-
-        7. Data processing:
-            - `--cleanup` (default: `False`): Clean old training files to start training from scratch.
-
-        8. Optimization:
-            - `--checkpointing` (default: `False`): Enable/disable checkpointing to save RAM.
-            - `--deterministic` (default: `False`): When enabled, uses deterministic algorithms to ensure consistent results for the same input data.
-            - `--benchmark` (default: `False`): When enabled, tests and selects the optimal algorithm for the hardware and specific size.
-    """)
-    quit()
-elif argv_is_allows[18] in argv:
-    print("""Usage:
-        1. `--help_audio_effects`: Help for adding audio effects.
-        2. `--help_audioldm2`: Help for music editing.
-        3. `--help_convert`: Help for audio conversion.
-        4. `--help_create_dataset`: Help for creating training data.
-        5. `--help_create_index`: Help for creating an index.
-        6. `--help_extract`: Help for extracting training data.
-        7. `--help_preprocess`: Help for preprocessing data.
-        8. `--help_separator_music`: Help for music separation.
-        9. `--help_train`: Help for model training.
-    """)
-    quit()
-
-if __name__ == "__main__":
-    if "--train" in argv:
-        import torch.multiprocessing as mp
-        mp.set_start_method("spawn")
-        
-    try:
-        main()
-    except:
-        pass

main/app/tabs/inference/inference.py

@@ -1,596 +0,0 @@
-import gradio as gr
-from main.tools import huggingface
-from main.configs.config import Config
-from main.app.based.utils import *
-
-def inference_tabs():
-    # Audio Conversion Tab
-    with gr.TabItem(translations["convert_audio"], visible=configs.get("convert_tab", True)):
-        gr.Markdown(f"## {translations['convert_audio']}")
-        with gr.Row():
-            gr.Markdown(translations["convert_info"])
-        
-        with gr.Row():
-            with gr.Column():
-                with gr.Accordion(translations["model_accordion"], open=True):
-                    with gr.Row(equal_height=True):
-                        model_pth = gr.Dropdown(label=translations["model_name"], choices=model_name, value=model_name[0] if len(model_name) >= 1 else "", interactive=True, allow_custom_value=True)
-                        model_index = gr.Dropdown(label=translations["index_path"], choices=index_path, value=index_path[0] if len(index_path) >= 1 else "", interactive=True, allow_custom_value=True)
-                        refesh = gr.Button(translations["refesh"])
-                
-        with gr.Row(): 
-            with gr.Column():
-                audio_select = gr.Dropdown(label=translations["select_separate"], choices=[], value="", interactive=True, allow_custom_value=True, visible=False)
-                convert_button_2 = gr.Button(translations["convert_audio"], visible=False)
-        with gr.Row(): 
-            with gr.Column():
-                input0 = gr.File(label=translations["drop_audio"], file_types=[".wav", ".mp3", ".flac", ".ogg", ".opus", ".m4a", ".mp4", ".aac", ".alac", ".wma", ".aiff", ".webm", ".ac3"])  
-                play_audio = gr.Audio(show_download_button=True, interactive=False, label=translations["input_audio"])
-        with gr.Row(): 
-            with gr.Column():
-                with gr.Row():
-                    index_strength = gr.Slider(label=translations["index_strength"], info=translations["index_strength_info"], minimum=0, maximum=1, value=0.5, step=0.01, interactive=True, visible=model_index.value != "")
-        with gr.Row(): 
-            with gr.Column():
-                with gr.Accordion(translations["input_output"], open=False):
-                    with gr.Column():
-                        export_format = gr.Radio(label=translations["export_format"], info=translations["export_info"], choices=["wav", "mp3", "flac", "ogg", "opus", "m4a", "mp4", "aac", "alac", "wma", "aiff", "webm", "ac3"], value="wav", interactive=True)
-                        input_audio0 = gr.Dropdown(label=translations["audio_path"], value="", choices=paths_for_files, info=translations["provide_audio"], allow_custom_value=True, interactive=True)
-                        output_audio = gr.Textbox(label=translations["output_path"], value="audios/output.wav", placeholder="audios/output.wav", info=translations["output_path_info"], interactive=True)
-                    with gr.Column():
-                        refesh0 = gr.Button(translations["refesh"])
-                with gr.Accordion(translations["setting"], open=False):
-                    with gr.Row():
-                        cleaner0 = gr.Checkbox(label=translations["clear_audio"], value=False, interactive=True)
-                        autotune = gr.Checkbox(label=translations["autotune"], value=False, interactive=True)
-                        use_audio = gr.Checkbox(label=translations["use_audio"], value=False, interactive=True)
-                        checkpointing = gr.Checkbox(label=translations["memory_efficient_training"], value=False, interactive=True)
-                    with gr.Row():
-                        use_original = gr.Checkbox(label=translations["convert_original"], value=False, interactive=True, visible=use_audio.value) 
-                        convert_backing = gr.Checkbox(label=translations["convert_backing"], value=False, interactive=True, visible=use_audio.value)   
-                        not_merge_backing = gr.Checkbox(label=translations["not_merge_backing"], value=False, interactive=True, visible=use_audio.value)
-                        merge_instrument = gr.Checkbox(label=translations["merge_instruments"], value=False, interactive=True, visible=use_audio.value) 
-                    with gr.Row():
-                        pitch = gr.Slider(minimum=-20, maximum=20, step=1, info=translations["pitch_info"], label=translations["pitch"], value=0, interactive=True)
-                        clean_strength0 = gr.Slider(label=translations["clean_strength"], info=translations["clean_strength_info"], minimum=0, maximum=1, value=0.5, step=0.1, interactive=True, visible=cleaner0.value)
-                    
-                    with gr.Accordion(translations["f0_method"], open=False):
-                        with gr.Group():
-                            with gr.Row():
-                                onnx_f0_mode = gr.Checkbox(label=translations["f0_onnx_mode"], info=translations["f0_onnx_mode_info"], value=False, interactive=True)
-                                unlock_full_method = gr.Checkbox(label=translations["f0_unlock"], info=translations["f0_unlock_info"], value=False, interactive=True)
-                            method = gr.Radio(label=translations["f0_method"], info=translations["f0_method_info"], choices=method_f0+["hybrid"], value="rmvpe", interactive=True)
-                            hybrid_method = gr.Dropdown(label=translations["f0_method_hybrid"], info=translations["f0_method_hybrid_info"], choices=["hybrid[pm+dio]", "hybrid[pm+crepe-tiny]", "hybrid[pm+crepe]", "hybrid[pm+fcpe]", "hybrid[pm+rmvpe]", "hybrid[pm+harvest]", "hybrid[pm+yin]", "hybrid[dio+crepe-tiny]", "hybrid[dio+crepe]", "hybrid[dio+fcpe]", "hybrid[dio+rmvpe]", "hybrid[dio+harvest]", "hybrid[dio+yin]", "hybrid[crepe-tiny+crepe]", "hybrid[crepe-tiny+fcpe]", "hybrid[crepe-tiny+rmvpe]", "hybrid[crepe-tiny+harvest]", "hybrid[crepe+fcpe]", "hybrid[crepe+rmvpe]", "hybrid[crepe+harvest]", "hybrid[crepe+yin]", "hybrid[fcpe+rmvpe]", "hybrid[fcpe+harvest]", "hybrid[fcpe+yin]", "hybrid[rmvpe+harvest]", "hybrid[rmvpe+yin]", "hybrid[harvest+yin]"], value="hybrid[pm+dio]", interactive=True, allow_custom_value=True, visible=method.value == "hybrid")
-                        hop_length = gr.Slider(label="Hop length", info=translations["hop_length_info"], minimum=1, maximum=512, value=128, step=1, interactive=True, visible=False)
-                    with gr.Accordion(translations["f0_file"], open=False):
-                        upload_f0_file = gr.File(label=translations["upload_f0"], file_types=[".txt"])  
-                        f0_file_dropdown = gr.Dropdown(label=translations["f0_file_2"], value="", choices=f0_file, allow_custom_value=True, interactive=True)
-                        refesh_f0_file = gr.Button(translations["refesh"])
-                    with gr.Accordion(translations["hubert_model"], open=False):
-                        embed_mode = gr.Radio(label=translations["embed_mode"], info=translations["embed_mode_info"], value="fairseq", choices=embedders_mode, interactive=True, visible=True)
-                        embedders = gr.Radio(label=translations["hubert_model"], info=translations["hubert_info"], choices=embedders_model, value="hubert_base", interactive=True)
-                        custom_embedders = gr.Textbox(label=translations["modelname"], info=translations["modelname_info"], value="", placeholder="hubert_base", interactive=True, visible=embedders.value == "custom")    
-                    with gr.Accordion(translations["use_presets"], open=False):
-                        with gr.Row():
-                            presets_name = gr.Dropdown(label=translations["file_preset"], choices=presets_file, value=presets_file[0] if len(presets_file) > 0 else '', interactive=True, allow_custom_value=True)
-                        with gr.Row():
-                            load_click = gr.Button(translations["load_file"], variant="primary")
-                            refesh_click = gr.Button(translations["refesh"])
-                        with gr.Accordion(translations["export_file"], open=False):
-                            with gr.Row():
-                                with gr.Column():
-                                    with gr.Group():
-                                        with gr.Row():
-                                            cleaner_chbox = gr.Checkbox(label=translations["save_clean"], value=True, interactive=True)
-                                            autotune_chbox = gr.Checkbox(label=translations["save_autotune"], value=True, interactive=True)
-                                            pitch_chbox = gr.Checkbox(label=translations["save_pitch"], value=True, interactive=True)
-                                            index_strength_chbox = gr.Checkbox(label=translations["save_index_2"], value=True, interactive=True)
-                                            resample_sr_chbox = gr.Checkbox(label=translations["save_resample"], value=True, interactive=True)
-                                            filter_radius_chbox = gr.Checkbox(label=translations["save_filter"], value=True, interactive=True)
-                                            volume_envelope_chbox = gr.Checkbox(label=translations["save_envelope"], value=True, interactive=True)
-                                            protect_chbox = gr.Checkbox(label=translations["save_protect"], value=True, interactive=True)
-                                            split_audio_chbox = gr.Checkbox(label=translations["save_split"], value=True, interactive=True)
-                                            formant_shifting_chbox = gr.Checkbox(label=translations["formantshift"], value=True, interactive=True)
-                                with gr.Row():
-                                    with gr.Column():
-                                        name_to_save_file = gr.Textbox(label=translations["filename_to_save"])
-                                        save_file_button = gr.Button(translations["export_file"])
-                        with gr.Row():
-                            upload_presets = gr.File(label=translations["upload_presets"], file_types=[".json"])  
-                    with gr.Column():
-                        with gr.Row():
-                            split_audio = gr.Checkbox(label=translations["split_audio"], value=False, interactive=True)
-                            formant_shifting = gr.Checkbox(label=translations["formantshift"], value=False, interactive=True)
-                        f0_autotune_strength = gr.Slider(minimum=0, maximum=1, label=translations["autotune_rate"], info=translations["autotune_rate_info"], value=1, step=0.1, interactive=True, visible=autotune.value)
-                        resample_sr = gr.Slider(minimum=0, maximum=96000, label=translations["resample"], info=translations["resample_info"], value=0, step=1, interactive=True)
-                        filter_radius = gr.Slider(minimum=0, maximum=7, label=translations["filter_radius"], info=translations["filter_radius_info"], value=3, step=1, interactive=True)
-                        volume_envelope = gr.Slider(minimum=0, maximum=1, label=translations["volume_envelope"], info=translations["volume_envelope_info"], value=1, step=0.1, interactive=True)
-                        protect = gr.Slider(minimum=0, maximum=1, label=translations["protect"], info=translations["protect_info"], value=0.5, step=0.01, interactive=True)
-                    with gr.Row():
-                        formant_qfrency = gr.Slider(value=1.0, label=translations["formant_qfrency"], info=translations["formant_qfrency"], minimum=0.0, maximum=16.0, step=0.1, interactive=True, visible=False)
-                        formant_timbre = gr.Slider(value=1.0, label=translations["formant_timbre"], info=translations["formant_timbre"], minimum=0.0, maximum=16.0, step=0.1, interactive=True, visible=False)
-        with gr.Row():
-            convert_button = gr.Button(translations["convert_audio"], variant="primary")
-        gr.Markdown(translations["output_convert"])
-        with gr.Row():
-            main_convert = gr.Audio(show_download_button=True, interactive=False, label=translations["main_convert"])
-            backing_convert = gr.Audio(show_download_button=True, interactive=False, label=translations["convert_backing"], visible=convert_backing.value)
-            main_backing = gr.Audio(show_download_button=True, interactive=False, label=translations["main_or_backing"], visible=convert_backing.value)  
-        with gr.Row():
-            original_convert = gr.Audio(show_download_button=True, interactive=False, label=translations["convert_original"], visible=use_original.value)
-            vocal_instrument = gr.Audio(show_download_button=True, interactive=False, label=translations["voice_or_instruments"], visible=merge_instrument.value)  
-        with gr.Row():
-            upload_f0_file.upload(fn=lambda inp: shutil.move(inp.name, os.path.join("assets", "f0")), inputs=[upload_f0_file], outputs=[f0_file_dropdown])
-            refesh_f0_file.click(fn=change_f0_choices, inputs=[], outputs=[f0_file_dropdown])
-            unlock_full_method.change(fn=unlock_f0, inputs=[unlock_full_method], outputs=[method])
-        with gr.Row():
-            load_click.click(
-                fn=load_presets, 
-                inputs=[
-                    presets_name, 
-                    cleaner0, 
-                    autotune, 
-                    pitch, 
-                    clean_strength0, 
-                    index_strength, 
-                    resample_sr, 
-                    filter_radius, 
-                    volume_envelope, 
-                    protect, 
-                    split_audio, 
-                    f0_autotune_strength, 
-                    formant_qfrency, 
-                    formant_timbre
-                ], 
-                outputs=[
-                    cleaner0, 
-                    autotune, 
-                    pitch, 
-                    clean_strength0, 
-                    index_strength, 
-                    resample_sr, 
-                    filter_radius, 
-                    volume_envelope, 
-                    protect, 
-                    split_audio, 
-                    f0_autotune_strength, 
-                    formant_shifting, 
-                    formant_qfrency, 
-                    formant_timbre
-                ]
-            )
-            refesh_click.click(fn=change_preset_choices, inputs=[], outputs=[presets_name])
-            save_file_button.click(
-                fn=save_presets, 
-                inputs=[
-                    name_to_save_file, 
-                    cleaner0, 
-                    autotune, 
-                    pitch, 
-                    clean_strength0, 
-                    index_strength, 
-                    resample_sr, 
-                    filter_radius, 
-                    volume_envelope, 
-                    protect, 
-                    split_audio, 
-                    f0_autotune_strength, 
-                    cleaner_chbox, 
-                    autotune_chbox, 
-                    pitch_chbox, 
-                    index_strength_chbox, 
-                    resample_sr_chbox, 
-                    filter_radius_chbox, 
-                    volume_envelope_chbox, 
-                    protect_chbox, 
-                    split_audio_chbox, 
-                    formant_shifting_chbox, 
-                    formant_shifting, 
-                    formant_qfrency, 
-                    formant_timbre
-                ], 
-                outputs=[presets_name]
-            )
-        with gr.Row():
-            upload_presets.upload(fn=lambda audio_in: shutil.move(audio_in.name, os.path.join("assets", "presets")), inputs=[upload_presets], outputs=[presets_name])
-            autotune.change(fn=visible, inputs=[autotune], outputs=[f0_autotune_strength])
-            use_audio.change(fn=lambda a: [visible(a), visible(a), visible(a), visible(a), visible(a), valueFalse_interactive(a), valueFalse_interactive(a), valueFalse_interactive(a), valueFalse_interactive(a), visible(not a), visible(not a), visible(not a), visible(not a)], inputs=[use_audio], outputs=[main_backing, use_original, convert_backing, not_merge_backing, merge_instrument, use_original, convert_backing, not_merge_backing, merge_instrument, input_audio0, output_audio, input0, play_audio])
-        with gr.Row():
-            convert_backing.change(fn=lambda a,b: [change_backing_choices(a, b), visible(a)], inputs=[convert_backing, not_merge_backing], outputs=[use_original, backing_convert])
-            use_original.change(fn=lambda audio, original: [visible(original), visible(not original), visible(audio and not original), valueFalse_interactive(not original), valueFalse_interactive(not original)], inputs=[use_audio, use_original], outputs=[original_convert, main_convert, main_backing, convert_backing, not_merge_backing])
-            cleaner0.change(fn=visible, inputs=[cleaner0], outputs=[clean_strength0])
-        with gr.Row():
-            merge_instrument.change(fn=visible, inputs=[merge_instrument], outputs=[vocal_instrument])
-            not_merge_backing.change(fn=lambda audio, merge, cvb: [visible(audio and not merge), change_backing_choices(cvb, merge)], inputs=[use_audio, not_merge_backing, convert_backing], outputs=[main_backing, use_original])
-            method.change(fn=lambda method, hybrid: [visible(method == "hybrid"), hoplength_show(method, hybrid)], inputs=[method, hybrid_method], outputs=[hybrid_method, hop_length])
-        with gr.Row():
-            hybrid_method.change(fn=hoplength_show, inputs=[method, hybrid_method], outputs=[hop_length])
-            refesh.click(fn=change_models_choices, inputs=[], outputs=[model_pth, model_index])
-            model_pth.change(fn=get_index, inputs=[model_pth], outputs=[model_index])
-        with gr.Row():
-            input0.upload(fn=lambda audio_in: shutil.move(audio_in.name, os.path.join("audios")), inputs=[input0], outputs=[input_audio0])
-            input_audio0.change(fn=lambda audio: audio if os.path.isfile(audio) else None, inputs=[input_audio0], outputs=[play_audio])
-            formant_shifting.change(fn=lambda a: [visible(a)]*2, inputs=[formant_shifting], outputs=[formant_qfrency, formant_timbre])
-        with gr.Row():
-            embedders.change(fn=lambda embedders: visible(embedders == "custom"), inputs=[embedders], outputs=[custom_embedders])
-            refesh0.click(fn=change_audios_choices, inputs=[input_audio0], outputs=[input_audio0])
-            model_index.change(fn=index_strength_show, inputs=[model_index], outputs=[index_strength])
-        with gr.Row():
-            audio_select.change(fn=lambda: visible(True), inputs=[], outputs=[convert_button_2])
-            convert_button.click(fn=lambda: visible(False), inputs=[], outputs=[convert_button])
-            convert_button_2.click(fn=lambda: [visible(False), visible(False)], inputs=[], outputs=[audio_select, convert_button_2])
-        with gr.Row():
-            convert_button.click(
-                fn=convert_selection,
-                inputs=[
-                    cleaner0,
-                    autotune,
-                    use_audio,
-                    use_original,
-                    convert_backing,
-                    not_merge_backing,
-                    merge_instrument,
-                    pitch,
-                    clean_strength0,
-                    model_pth,
-                    model_index,
-                    index_strength,
-                    input_audio0,
-                    output_audio,
-                    export_format,
-                    method,
-                    hybrid_method,
-                    hop_length,
-                    embedders,
-                    custom_embedders,
-                    resample_sr,
-                    filter_radius,
-                    volume_envelope,
-                    protect,
-                    split_audio,
-                    f0_autotune_strength,
-                    checkpointing,
-                    onnx_f0_mode,
-                    formant_shifting, 
-                    formant_qfrency, 
-                    formant_timbre,
-                    f0_file_dropdown,
-                    embed_mode
-                ],
-                outputs=[audio_select, main_convert, backing_convert, main_backing, original_convert, vocal_instrument, convert_button],
-                api_name="convert_selection"
-            )
-            embed_mode.change(fn=visible_embedders, inputs=[embed_mode], outputs=[embedders])
-            convert_button_2.click(
-                fn=convert_audio,
-                inputs=[
-                    cleaner0,
-                    autotune,
-                    use_audio,
-                    use_original,
-                    convert_backing,
-                    not_merge_backing,
-                    merge_instrument,
-                    pitch,
-                    clean_strength0,
-                    model_pth,
-                    model_index,
-                    index_strength,
-                    input_audio0,
-                    output_audio,
-                    export_format,
-                    method,
-                    hybrid_method,
-                    hop_length,
-                    embedders,
-                    custom_embedders,
-                    resample_sr,
-                    filter_radius,
-                    volume_envelope,
-                    protect,
-                    split_audio,
-                    f0_autotune_strength,
-                    audio_select,
-                    checkpointing,
-                    onnx_f0_mode,
-                    formant_shifting, 
-                    formant_qfrency, 
-                    formant_timbre,
-                    f0_file_dropdown,
-                    embed_mode
-                ],
-                outputs=[main_convert, backing_convert, main_backing, original_convert, vocal_instrument, convert_button],
-                api_name="convert_audio"
-            )
-
-    # Text-to-Speech Conversion Tab
-    with gr.TabItem(translations["convert_text"], visible=configs.get("tts_tab", True)):
-        gr.Markdown(translations["convert_text_markdown"])
-        with gr.Row():
-            gr.Markdown(translations["convert_text_markdown_2"])
-        with gr.Accordion(translations["model_accordion"], open=True):
-            with gr.Row(equal_height=True):
-                model_pth0 = gr.Dropdown(label=translations["model_name"], choices=model_name, value=model_name[0] if len(model_name) >= 1 else "", interactive=True, allow_custom_value=True)
-                model_index0 = gr.Dropdown(label=translations["index_path"], choices=index_path, value=index_path[0] if len(index_path) >= 1 else "", interactive=True, allow_custom_value=True)
-                refesh1 = gr.Button(translations["refesh"])
-                    
-        with gr.Row():
-            with gr.Column():
-                with gr.Group():
-                    with gr.Row():
-                        use_txt = gr.Checkbox(label=translations["input_txt"], value=False, interactive=True)
-                        google_tts_check_box = gr.Checkbox(label=translations["googletts"], value=False, interactive=True)
-                    prompt = gr.Textbox(label=translations["text_to_speech"], value="", placeholder="Hello Words", lines=3)
-            with gr.Column():
-                speed = gr.Slider(label=translations["voice_speed"], info=translations["voice_speed_info"], minimum=-100, maximum=100, value=0, step=1)
-                pitch0 = gr.Slider(minimum=-20, maximum=20, step=1, info=translations["pitch_info"], label=translations["pitch"], value=0, interactive=True)
-        with gr.Row():
-            tts_button = gr.Button(translations["tts_1"], variant="primary", scale=2)
-            convert_button0 = gr.Button(translations["tts_2"], variant="secondary", scale=2)
-        with gr.Row():
-            with gr.Column():
-                txt_input = gr.File(label=translations["drop_text"], file_types=[".txt", ".srt"], visible=use_txt.value)  
-                tts_voice = gr.Dropdown(label=translations["voice"], choices=edgetts, interactive=True, value="vi-VN-NamMinhNeural")
-                tts_pitch = gr.Slider(minimum=-20, maximum=20, step=1, info=translations["pitch_info_2"], label=translations["pitch"], value=0, interactive=True)
-            with gr.Column():
-                
-                with gr.Row():
-                    index_strength0 = gr.Slider(label=translations["index_strength"], info=translations["index_strength_info"], minimum=0, maximum=1, value=0.5, step=0.01, interactive=True, visible=model_index0.value != "")
-                with gr.Accordion(translations["output_path"], open=False):
-                    export_format0 = gr.Radio(label=translations["export_format"], info=translations["export_info"], choices=["wav", "mp3", "flac", "ogg", "opus", "m4a", "mp4", "aac", "alac", "wma", "aiff", "webm", "ac3"], value="wav", interactive=True)
-                    output_audio0 = gr.Textbox(label=translations["output_tts"], value="audios/tts.wav", placeholder="audios/tts.wav", info=translations["tts_output"], interactive=True)
-                    output_audio1 = gr.Textbox(label=translations["output_tts_convert"], value="audios/tts-convert.wav", placeholder="audios/tts-convert.wav", info=translations["tts_output"], interactive=True)
-                with gr.Accordion(translations["setting"], open=False):
-                    with gr.Accordion(translations["f0_method"], open=False):
-                        with gr.Group():
-                            with gr.Row():
-                                onnx_f0_mode1 = gr.Checkbox(label=translations["f0_onnx_mode"], info=translations["f0_onnx_mode_info"], value=False, interactive=True)
-                                unlock_full_method3 = gr.Checkbox(label=translations["f0_unlock"], info=translations["f0_unlock_info"], value=False, interactive=True)
-                            method0 = gr.Radio(label=translations["f0_method"], info=translations["f0_method_info"], choices=method_f0+["hybrid"], value="rmvpe", interactive=True)
-                            hybrid_method0 = gr.Dropdown(label=translations["f0_method_hybrid"], info=translations["f0_method_hybrid_info"], choices=["hybrid[pm+dio]", "hybrid[pm+crepe-tiny]", "hybrid[pm+crepe]", "hybrid[pm+fcpe]", "hybrid[pm+rmvpe]", "hybrid[pm+harvest]", "hybrid[pm+yin]", "hybrid[dio+crepe-tiny]", "hybrid[dio+crepe]", "hybrid[dio+fcpe]", "hybrid[dio+rmvpe]", "hybrid[dio+harvest]", "hybrid[dio+yin]", "hybrid[crepe-tiny+crepe]", "hybrid[crepe-tiny+fcpe]", "hybrid[crepe-tiny+rmvpe]", "hybrid[crepe-tiny+harvest]", "hybrid[crepe+fcpe]", "hybrid[crepe+rmvpe]", "hybrid[crepe+harvest]", "hybrid[crepe+yin]", "hybrid[fcpe+rmvpe]", "hybrid[fcpe+harvest]", "hybrid[fcpe+yin]", "hybrid[rmvpe+harvest]", "hybrid[rmvpe+yin]", "hybrid[harvest+yin]"], value="hybrid[pm+dio]", interactive=True, allow_custom_value=True, visible=method0.value == "hybrid")
-                        hop_length0 = gr.Slider(label="Hop length", info=translations["hop_length_info"], minimum=1, maximum=512, value=128, step=1, interactive=True, visible=False)
-                    with gr.Accordion(translations["f0_file"], open=False):
-                        upload_f0_file0 = gr.File(label=translations["upload_f0"], file_types=[".txt"])  
-                        f0_file_dropdown0 = gr.Dropdown(label=translations["f0_file_2"], value="", choices=f0_file, allow_custom_value=True, interactive=True)
-                        refesh_f0_file0 = gr.Button(translations["refesh"])
-                    with gr.Accordion(translations["hubert_model"], open=False):
-                        embed_mode1 = gr.Radio(label=translations["embed_mode"], info=translations["embed_mode_info"], value="fairseq", choices=embedders_mode, interactive=True, visible=True)
-                        embedders0 = gr.Radio(label=translations["hubert_model"], info=translations["hubert_info"], choices=embedders_model, value="hubert_base", interactive=True)
-                        custom_embedders0 = gr.Textbox(label=translations["modelname"], info=translations["modelname_info"], value="", placeholder="hubert_base", interactive=True, visible=embedders0.value == "custom")
-                    with gr.Group():
-                        with gr.Row():
-                            formant_shifting1 = gr.Checkbox(label=translations["formantshift"], value=False, interactive=True)  
-                            split_audio0 = gr.Checkbox(label=translations["split_audio"], value=False, interactive=True)   
-                            cleaner1 = gr.Checkbox(label=translations["clear_audio"], value=False, interactive=True)     
-                            autotune3 = gr.Checkbox(label=translations["autotune"], value=False, interactive=True) 
-                            checkpointing0 = gr.Checkbox(label=translations["memory_efficient_training"], value=False, interactive=True)         
-                    with gr.Column():
-                        f0_autotune_strength0 = gr.Slider(minimum=0, maximum=1, label=translations["autotune_rate"], info=translations["autotune_rate_info"], value=1, step=0.1, interactive=True, visible=autotune3.value)
-                        clean_strength1 = gr.Slider(label=translations["clean_strength"], info=translations["clean_strength_info"], minimum=0, maximum=1, value=0.5, step=0.1, interactive=True, visible=cleaner1.value)
-                        resample_sr0 = gr.Slider(minimum=0, maximum=96000, label=translations["resample"], info=translations["resample_info"], value=0, step=1, interactive=True)
-                        filter_radius0 = gr.Slider(minimum=0, maximum=7, label=translations["filter_radius"], info=translations["filter_radius_info"], value=3, step=1, interactive=True)
-                        volume_envelope0 = gr.Slider(minimum=0, maximum=1, label=translations["volume_envelope"], info=translations["volume_envelope_info"], value=1, step=0.1, interactive=True)
-                        protect0 = gr.Slider(minimum=0, maximum=1, label=translations["protect"], info=translations["protect_info"], value=0.5, step=0.01, interactive=True)
-                    with gr.Row():
-                        formant_qfrency1 = gr.Slider(value=1.0, label=translations["formant_qfrency"], info=translations["formant_qfrency"], minimum=0.0, maximum=16.0, step=0.1, interactive=True, visible=False)
-                        formant_timbre1 = gr.Slider(value=1.0, label=translations["formant_timbre"], info=translations["formant_timbre"], minimum=0.0, maximum=16.0, step=0.1, interactive=True, visible=False)
-        with gr.Row():
-            gr.Markdown(translations["output_tts_markdown"])
-        with gr.Row():
-            tts_voice_audio = gr.Audio(show_download_button=True, interactive=False, label=translations["output_text_to_speech"])
-            tts_voice_convert = gr.Audio(show_download_button=True, interactive=False, label=translations["output_file_tts_convert"])
-        with gr.Row():
-            unlock_full_method3.change(fn=unlock_f0, inputs=[unlock_full_method3], outputs=[method0])
-            upload_f0_file0.upload(fn=lambda inp: shutil.move(inp.name, os.path.join("assets", "f0")), inputs=[upload_f0_file0], outputs=[f0_file_dropdown0])
-            refesh_f0_file0.click(fn=change_f0_choices, inputs=[], outputs=[f0_file_dropdown0])
-        with gr.Row():
-            embed_mode1.change(fn=visible_embedders, inputs=[embed_mode1], outputs=[embedders0])
-            autotune3.change(fn=visible, inputs=[autotune3], outputs=[f0_autotune_strength0])
-            model_pth0.change(fn=get_index, inputs=[model_pth0], outputs=[model_index0])
-        with gr.Row():
-            cleaner1.change(fn=visible, inputs=[cleaner1], outputs=[clean_strength1])
-            method0.change(fn=lambda method, hybrid: [visible(method == "hybrid"), hoplength_show(method, hybrid)], inputs=[method0, hybrid_method0], outputs=[hybrid_method0, hop_length0])
-            hybrid_method0.change(fn=hoplength_show, inputs=[method0, hybrid_method0], outputs=[hop_length0])
-        with gr.Row():
-            refesh1.click(fn=change_models_choices, inputs=[], outputs=[model_pth0, model_index0])
-            embedders0.change(fn=lambda embedders: visible(embedders == "custom"), inputs=[embedders0], outputs=[custom_embedders0])
-            formant_shifting1.change(fn=lambda a: [visible(a)]*2, inputs=[formant_shifting1], outputs=[formant_qfrency1, formant_timbre1])
-        with gr.Row():
-            model_index0.change(fn=index_strength_show, inputs=[model_index0], outputs=[index_strength0])
-            txt_input.upload(fn=process_input, inputs=[txt_input], outputs=[prompt])
-            use_txt.change(fn=visible, inputs=[use_txt], outputs=[txt_input])
-        with gr.Row():
-            google_tts_check_box.change(fn=change_tts_voice_choices, inputs=[google_tts_check_box], outputs=[tts_voice])
-            tts_button.click(
-                fn=TTS, 
-                inputs=[
-                    prompt, 
-                    tts_voice, 
-                    speed, 
-                    output_audio0,
-                    tts_pitch,
-                    google_tts_check_box,
-                    txt_input
-                ], 
-                outputs=[tts_voice_audio],
-                api_name="text-to-speech"
-            )
-            convert_button0.click(
-                fn=convert_tts,
-                inputs=[
-                    cleaner1, 
-                    autotune3, 
-                    pitch0, 
-                    clean_strength1, 
-                    model_pth0, 
-                    model_index0, 
-                    index_strength0, 
-                    output_audio0, 
-                    output_audio1,
-                    export_format0,
-                    method0, 
-                    hybrid_method0, 
-                    hop_length0, 
-                    embedders0, 
-                    custom_embedders0, 
-                    resample_sr0, 
-                    filter_radius0, 
-                    volume_envelope0, 
-                    protect0,
-                    split_audio0,
-                    f0_autotune_strength0,
-                    checkpointing0,
-                    onnx_f0_mode1,
-                    formant_shifting1, 
-                    formant_qfrency1, 
-                    formant_timbre1,
-                    f0_file_dropdown0,
-                    embed_mode1
-                ],
-                outputs=[tts_voice_convert],
-                api_name="convert_tts"
-            )
-
-    # Whisper Conversion Tab
-    with gr.TabItem(translations["convert_with_whisper"], visible=configs.get("convert_with_whisper", True)):
-        gr.Markdown(f"## {translations['convert_with_whisper']}")
-        with gr.Row():
-            gr.Markdown(translations["convert_with_whisper_info"])
-        with gr.Row():
-            with gr.Column():
-                with gr.Accordion(translations["model_accordion"] + " 1", open=True):
-                    with gr.Row(equal_height=True):
-                        model_pth2 = gr.Dropdown(label=translations["model_name"], choices=model_name, value=model_name[0] if len(model_name) >= 1 else "", interactive=True, allow_custom_value=True)
-                        model_index2 = gr.Dropdown(label=translations["index_path"], choices=index_path, value=index_path[0] if len(index_path) >= 1 else "", interactive=True, allow_custom_value=True)
-                        refesh2 = gr.Button(translations["refesh"])
-                with gr.Accordion(translations["model_accordion"] + " 2", open=True):
-                    with gr.Row(equal_height=True):
-                        model_pth3 = gr.Dropdown(label=translations["model_name"], choices=model_name, value=model_name[0] if len(model_name) >= 1 else "", interactive=True, allow_custom_value=True)
-                        model_index3 = gr.Dropdown(label=translations["index_path"], choices=index_path, value=index_path[0] if len(index_path) >= 1 else "", interactive=True, allow_custom_value=True)
-                        refesh3 = gr.Button(translations["refesh"])
-                with gr.Group():
-                    with gr.Row():
-                        cleaner2 = gr.Checkbox(label=translations["clear_audio"], value=False, interactive=True)
-                        autotune2 = gr.Checkbox(label=translations["autotune"], value=False, interactive=True)
-                        checkpointing2 = gr.Checkbox(label=translations["memory_efficient_training"], value=False, interactive=True)
-                        formant_shifting2 = gr.Checkbox(label=translations["formantshift"], value=False, interactive=True)
-                    with gr.Row():
-                        num_spk = gr.Slider(minimum=2, maximum=8, step=1, info=translations["num_spk_info"], label=translations["num_spk"], value=2, interactive=True)
-        with gr.Row():
-            with gr.Column():
-                convert_button3 = gr.Button(translations["convert_audio"], variant="primary")
-        with gr.Row():
-            with gr.Column():
-                with gr.Row():
-                    pitch3 = gr.Slider(minimum=-20, maximum=20, step=1, info=translations["pitch_info"], label=translations["pitch"], value=0, interactive=True)
-                    index_strength2 = gr.Slider(label=translations["index_strength"], info=translations["index_strength_info"], minimum=0, maximum=1, value=0.5, step=0.01, interactive=True, visible=model_index2.value != "")
-                with gr.Accordion(translations["input_output"], open=False):
-                    with gr.Column():
-                        export_format2 = gr.Radio(label=translations["export_format"], info=translations["export_info"], choices=["wav", "mp3", "flac", "ogg", "opus", "m4a", "mp4", "aac", "alac", "wma", "aiff", "webm", "ac3"], value="wav", interactive=True)
-                        input_audio1 = gr.Dropdown(label=translations["audio_path"], value="", choices=paths_for_files, info=translations["provide_audio"], allow_custom_value=True, interactive=True)
-                        output_audio2 = gr.Textbox(label=translations["output_path"], value="audios/output.wav", placeholder="audios/output.wav", info=translations["output_path_info"], interactive=True)
-                    with gr.Column():
-                        refesh4 = gr.Button(translations["refesh"])
-                    with gr.Row():
-                        input2 = gr.File(label=translations["drop_audio"], file_types=[".wav", ".mp3", ".flac", ".ogg", ".opus", ".m4a", ".mp4", ".aac", ".alac", ".wma", ".aiff", ".webm", ".ac3"])
-                with gr.Column():
-                    with gr.Row():
-                        pitch4 = gr.Slider(minimum=-20, maximum=20, step=1, info=translations["pitch_info"], label=translations["pitch"], value=0, interactive=True)
-                        index_strength3 = gr.Slider(label=translations["index_strength"], info=translations["index_strength_info"], minimum=0, maximum=1, value=0.5, step=0.01, interactive=True, visible=model_index3.value != "")
-                with gr.Accordion(translations["setting"], open=False):
-                    with gr.Row():
-                        model_size = gr.Radio(label=translations["model_size"], info=translations["model_size_info"], choices=["tiny", "tiny.en", "base", "base.en", "small", "small.en", "medium", "medium.en", "large-v1", "large-v2", "large-v3", "large-v3-turbo"], value="medium", interactive=True)
-                    with gr.Accordion(translations["f0_method"], open=False):
-                        with gr.Group():
-                            with gr.Row():
-                                onnx_f0_mode4 = gr.Checkbox(label=translations["f0_onnx_mode"], info=translations["f0_onnx_mode_info"], value=False, interactive=True)
-                                unlock_full_method2 = gr.Checkbox(label=translations["f0_unlock"], info=translations["f0_unlock_info"], value=False, interactive=True)
-                            method3 = gr.Radio(label=translations["f0_method"], info=translations["f0_method_info"], choices=method_f0+["hybrid"], value="rmvpe", interactive=True)
-                            hybrid_method3 = gr.Dropdown(label=translations["f0_method_hybrid"], info=translations["f0_method_hybrid_info"], choices=["hybrid[pm+dio]", "hybrid[pm+crepe-tiny]", "hybrid[pm+crepe]", "hybrid[pm+fcpe]", "hybrid[pm+rmvpe]", "hybrid[pm+harvest]", "hybrid[pm+yin]", "hybrid[dio+crepe-tiny]", "hybrid[dio+crepe]", "hybrid[dio+fcpe]", "hybrid[dio+rmvpe]", "hybrid[dio+harvest]", "hybrid[dio+yin]", "hybrid[crepe-tiny+crepe]", "hybrid[crepe-tiny+fcpe]", "hybrid[crepe-tiny+rmvpe]", "hybrid[crepe-tiny+harvest]", "hybrid[crepe+fcpe]", "hybrid[crepe+rmvpe]", "hybrid[crepe+harvest]", "hybrid[crepe+yin]", "hybrid[fcpe+rmvpe]", "hybrid[fcpe+harvest]", "hybrid[fcpe+yin]", "hybrid[rmvpe+harvest]", "hybrid[rmvpe+yin]", "hybrid[harvest+yin]"], value="hybrid[pm+dio]", interactive=True, allow_custom_value=True, visible=method3.value == "hybrid")
-                        hop_length3 = gr.Slider(label="Hop length", info=translations["hop_length_info"], minimum=1, maximum=512, value=128, step=1, interactive=True, visible=False)
-                    with gr.Accordion(translations["hubert_model"], open=False):
-                        embed_mode3 = gr.Radio(label=translations["embed_mode"], info=translations["embed_mode_info"], value="fairseq", choices=embedders_mode, interactive=True, visible=True)
-                        embedders3 = gr.Radio(label=translations["hubert_model"], info=translations["hubert_info"], choices=embedders_model, value="hubert_base", interactive=True)
-                        custom_embedders3 = gr.Textbox(label=translations["modelname"], info=translations["modelname_info"], value="", placeholder="hubert_base", interactive=True, visible=embedders3.value == "custom")
-                    with gr.Column():      
-                        clean_strength3 = gr.Slider(label=translations["clean_strength"], info=translations["clean_strength_info"], minimum=0, maximum=1, value=0.5, step=0.1, interactive=True, visible=cleaner2.value)
-                        f0_autotune_strength3 = gr.Slider(minimum=0, maximum=1, label=translations["autotune_rate"], info=translations["autotune_rate_info"], value=1, step=0.1, interactive=True, visible=autotune.value)
-                        resample_sr3 = gr.Slider(minimum=0, maximum=96000, label=translations["resample"], info=translations["resample_info"], value=0, step=1, interactive=True)
-                        filter_radius3 = gr.Slider(minimum=0, maximum=7, label=translations["filter_radius"], info=translations["filter_radius_info"], value=3, step=1, interactive=True)
-                        volume_envelope3 = gr.Slider(minimum=0, maximum=1, label=translations["volume_envelope"], info=translations["volume_envelope_info"], value=1, step=0.1, interactive=True)
-                        protect3 = gr.Slider(minimum=0, maximum=1, label=translations["protect"], info=translations["protect_info"], value=0.5, step=0.01, interactive=True)
-                    with gr.Row():
-                        formant_qfrency3 = gr.Slider(value=1.0, label=translations["formant_qfrency"] + " 1", info=translations["formant_qfrency"], minimum=0.0, maximum=16.0, step=0.1, interactive=True, visible=False)
-                        formant_timbre3 = gr.Slider(value=1.0, label=translations["formant_timbre"] + " 1", info=translations["formant_timbre"], minimum=0.0, maximum=16.0, step=0.1, interactive=True, visible=False)
-                    with gr.Row():
-                        formant_qfrency4 = gr.Slider(value=1.0, label=translations["formant_qfrency"] + " 2", info=translations["formant_qfrency"], minimum=0.0, maximum=16.0, step=0.1, interactive=True, visible=False)
-                        formant_timbre4 = gr.Slider(value=1.0, label=translations["formant_timbre"] + " 2", info=translations["formant_timbre"], minimum=0.0, maximum=16.0, step=0.1, interactive=True, visible=False)
-        with gr.Row():
-            gr.Markdown(translations["input_output"])
-        with gr.Row():
-            play_audio2 = gr.Audio(show_download_button=True, interactive=False, label=translations["input_audio"])
-            play_audio3 = gr.Audio(show_download_button=True, interactive=False, label=translations["output_file_tts_convert"])
-        with gr.Row():
-            autotune2.change(fn=visible, inputs=[autotune2], outputs=[f0_autotune_strength3])
-            cleaner2.change(fn=visible, inputs=[cleaner2], outputs=[clean_strength3])
-            method3.change(fn=lambda method, hybrid: [visible(method == "hybrid"), hoplength_show(method, hybrid)], inputs=[method3, hybrid_method3], outputs=[hybrid_method3, hop_length3])
-        with gr.Row():
-            hybrid_method3.change(fn=hoplength_show, inputs=[method3, hybrid_method3], outputs=[hop_length3])
-            refesh2.click(fn=change_models_choices, inputs=[], outputs=[model_pth2, model_index2])
-            model_pth2.change(fn=get_index, inputs=[model_pth2], outputs=[model_index2])
-        with gr.Row():
-            refesh3.click(fn=change_models_choices, inputs=[], outputs=[model_pth3, model_index3])
-            model_pth3.change(fn=get_index, inputs=[model_pth3], outputs=[model_index3])
-            input2.upload(fn=lambda audio_in: shutil.move(audio_in.name, os.path.join("audios")), inputs=[input2], outputs=[input_audio1])
-        with gr.Row():
-            input_audio1.change(fn=lambda audio: audio if os.path.isfile(audio) else None, inputs=[input_audio1], outputs=[play_audio2])
-            formant_shifting2.change(fn=lambda a: [visible(a)]*4, inputs=[formant_shifting2], outputs=[formant_qfrency3, formant_timbre3, formant_qfrency4, formant_timbre4])
-            embedders3.change(fn=lambda embedders: visible(embedders == "custom"), inputs=[embedders3], outputs=[custom_embedders3])
-        with gr.Row():
-            refesh4.click(fn=change_audios_choices, inputs=[input_audio1], outputs=[input_audio1])
-            model_index2.change(fn=index_strength_show, inputs=[model_index2], outputs=[index_strength2])
-            model_index3.change(fn=index_strength_show, inputs=[model_index3], outputs=[index_strength3])
-        with gr.Row():
-            unlock_full_method2.change(fn=unlock_f0, inputs=[unlock_full_method2], outputs=[method3])
-            embed_mode3.change(fn=visible_embedders, inputs=[embed_mode3], outputs=[embedders3])
-            convert_button3.click(
-                fn=convert_with_whisper,
-                inputs=[
-                    num_spk,
-                    model_size,
-                    cleaner2,
-                    clean_strength3,
-                    autotune2,
-                    f0_autotune_strength3,
-                    checkpointing2,
-                    model_pth2,
-                    model_pth3,
-                    model_index2,
-                    model_index3,
-                    pitch3,
-                    pitch4,
-                    index_strength2,
-                    index_strength3,
-                    export_format2,
-                    input_audio1,
-                    output_audio2,
-                    onnx_f0_mode4,
-                    method3,
-                    hybrid_method3,
-                    hop_length3,
-                    embed_mode3,
-                    embedders3,
-                    custom_embedders3,
-                    resample_sr3,
-                    filter_radius3,
-                    volume_envelope3,
-                    protect3,
-                    formant_shifting2,
-                    formant_qfrency3,
-                    formant_timbre3,
-                    formant_qfrency4,
-                    formant_timbre4,
-                ],
-                outputs=[play_audio3],
-                api_name="convert_with_whisper"
-            )

main/app/tabs/models/model.py

@@ -1,465 +0,0 @@
-from main.tools import huggingface
-from main.configs.config import Config
-from main.app.based.utils import *
-import gradio as gr
-
-
-def model_tabs():
-    with gr.Tabs():
-        with gr.Tab(label=translations["downloads"], visible=configs.get("downloads_tab", True)):
-            gr.Markdown(translations["download_markdown"])
-            with gr.Row():
-                gr.Markdown(translations["download_markdown_2"])
-            with gr.Row():
-                with gr.Accordion(translations["model_download"], open=True):
-                    with gr.Row():
-                        downloadmodel = gr.Radio(label=translations["model_download_select"], choices=[translations["download_url"], translations["download_from_csv"], translations["search_models"], translations["upload"]], interactive=True, value=translations["download_url"])
-                    with gr.Row():
-                        gr.Markdown("___")
-                    with gr.Column():
-                        with gr.Row():
-                            url_input = gr.Textbox(label=translations["model_url"], value="", placeholder="https://...", scale=6)
-                            download_model_name = gr.Textbox(label=translations["modelname"], value="", placeholder=translations["modelname"], scale=2)
-                        url_download = gr.Button(value=translations["downloads"], scale=2)
-                    with gr.Column():
-                        model_browser = gr.Dropdown(choices=models.keys(), label=translations["model_warehouse"], scale=8, allow_custom_value=True, visible=False)
-                        download_from_browser = gr.Button(value=translations["get_model"], scale=2, variant="primary", visible=False)
-                    with gr.Column():
-                        search_name = gr.Textbox(label=translations["name_to_search"], placeholder=translations["modelname"], interactive=True, scale=8, visible=False)
-                        search = gr.Button(translations["search_2"], scale=2, visible=False)
-                        search_dropdown = gr.Dropdown(label=translations["select_download_model"], value="", choices=[], allow_custom_value=True, interactive=False, visible=False)
-                        download = gr.Button(translations["downloads"], variant="primary", visible=False)
-                    with gr.Column():
-                        model_upload = gr.File(label=translations["drop_model"], file_types=[".pth", ".onnx", ".index", ".zip"], visible=False)
-            with gr.Row():
-                with gr.Accordion(translations["download_pretrained_2"], open=False):
-                    with gr.Row():
-                        pretrain_download_choices = gr.Radio(label=translations["model_download_select"], choices=[translations["download_url"], translations["list_model"], translations["upload"]], value=translations["download_url"], interactive=True)  
-                    with gr.Row():
-                        gr.Markdown("___")
-                    with gr.Column():
-                        with gr.Row():
-                            pretrainD = gr.Textbox(label=translations["pretrained_url"].format(dg="D"), value="", info=translations["only_huggingface"], placeholder="https://...", interactive=True, scale=4)
-                            pretrainG = gr.Textbox(label=translations["pretrained_url"].format(dg="G"), value="", info=translations["only_huggingface"], placeholder="https://...", interactive=True, scale=4)
-                        download_pretrain_button = gr.Button(translations["downloads"], scale=2)
-                    with gr.Column():
-                        with gr.Row():
-                            pretrain_choices = gr.Dropdown(label=translations["select_pretrain"], info=translations["select_pretrain_info"], choices=list(fetch_pretrained_data().keys()), value="Titan_Medium", allow_custom_value=True, interactive=True, scale=6, visible=False)
-                            sample_rate_pretrain = gr.Dropdown(label=translations["pretrain_sr"], info=translations["pretrain_sr"], choices=["48k", "40k", "32k"], value="48k", interactive=True, visible=False)
-                        download_pretrain_choices_button = gr.Button(translations["downloads"], scale=2, variant="primary", visible=False)
-                    with gr.Row():
-                        pretrain_upload_g = gr.File(label=translations["drop_pretrain"].format(dg="G"), file_types=[".pth"], visible=False)
-                        pretrain_upload_d = gr.File(label=translations["drop_pretrain"].format(dg="D"), file_types=[".pth"], visible=False)
-            with gr.Row():
-                url_download.click(
-                    fn=download_model, 
-                    inputs=[
-                        url_input, 
-                        download_model_name
-                    ], 
-                    outputs=[url_input],
-                    api_name="download_model"
-                )
-                download_from_browser.click(
-                    fn=lambda model: download_model(models[model], model), 
-                    inputs=[model_browser], 
-                    outputs=[model_browser],
-                    api_name="download_browser"
-                )
-            with gr.Row():
-                downloadmodel.change(fn=change_download_choices, inputs=[downloadmodel], outputs=[url_input, download_model_name, url_download, model_browser, download_from_browser, search_name, search, search_dropdown, download, model_upload])
-                search.click(fn=search_models, inputs=[search_name], outputs=[search_dropdown, download])
-                model_upload.upload(fn=save_drop_model, inputs=[model_upload], outputs=[model_upload])
-                download.click(
-                    fn=lambda model: download_model(model_options[model], model), 
-                    inputs=[search_dropdown], 
-                    outputs=[search_dropdown],
-                    api_name="search_models"
-                )
-            with gr.Row():
-                pretrain_download_choices.change(fn=change_download_pretrained_choices, inputs=[pretrain_download_choices], outputs=[pretrainD, pretrainG, download_pretrain_button, pretrain_choices, sample_rate_pretrain, download_pretrain_choices_button, pretrain_upload_d, pretrain_upload_g])
-                pretrain_choices.change(fn=update_sample_rate_dropdown, inputs=[pretrain_choices], outputs=[sample_rate_pretrain])
-            with gr.Row():
-                download_pretrain_button.click(
-                    fn=download_pretrained_model,
-                    inputs=[
-                        pretrain_download_choices, 
-                        pretrainD, 
-                        pretrainG
-                    ],
-                    outputs=[pretrainD],
-                    api_name="download_pretrain_link"
-                )
-                download_pretrain_choices_button.click(
-                    fn=download_pretrained_model,
-                    inputs=[
-                        pretrain_download_choices, 
-                        pretrain_choices, 
-                        sample_rate_pretrain
-                    ],
-                    outputs=[pretrain_choices],
-                    api_name="download_pretrain_choices"
-                )
-                pretrain_upload_g.upload(
-                    fn=lambda pretrain_upload_g: shutil.move(pretrain_upload_g.name, os.path.join("assets", "models", "pretrained_custom")), 
-                    inputs=[pretrain_upload_g], 
-                    outputs=[],
-                    api_name="upload_pretrain_g"
-                )
-                pretrain_upload_d.upload(
-                    fn=lambda pretrain_upload_d: shutil.move(pretrain_upload_d.name, os.path.join("assets", "models", "pretrained_custom")), 
-                    inputs=[pretrain_upload_d], 
-                    outputs=[],
-                    api_name="upload_pretrain_d"
-                )
-
-        with gr.Tab(label=translations["createdataset"], visible=configs.get("create_dataset_tab", True)):
-            gr.Markdown(translations["create_dataset_markdown"])
-            with gr.Row():
-                gr.Markdown(translations["create_dataset_markdown_2"])
-            with gr.Row():
-                dataset_url = gr.Textbox(label=translations["url_audio"], info=translations["create_dataset_url"], value="", placeholder="https://www.youtube.com/...", interactive=True)
-                output_dataset = gr.Textbox(label=translations["output_data"], info=translations["output_data_info"], value="dataset", placeholder="dataset", interactive=True)
-            with gr.Row():
-                with gr.Column():
-                    with gr.Group():
-                        with gr.Row():
-                            separator_reverb = gr.Checkbox(label=translations["dereveb_audio"], value=False, interactive=True)
-                            denoise_mdx = gr.Checkbox(label=translations["denoise"], value=False, interactive=True)
-                        with gr.Row():
-                            kim_vocal_version = gr.Radio(label=translations["model_ver"], info=translations["model_ver_info"], choices=["Version-1", "Version-2"], value="Version-2", interactive=True)
-                            kim_vocal_overlap = gr.Radio(label=translations["overlap"], info=translations["overlap_info"], choices=["0.25", "0.5", "0.75", "0.99"], value="0.25", interactive=True)
-                        with gr.Row():    
-                            kim_vocal_hop_length = gr.Slider(label="Hop length", info=translations["hop_length_info"], minimum=1, maximum=8192, value=1024, step=1, interactive=True)
-                            kim_vocal_batch_size = gr.Slider(label=translations["batch_size"], info=translations["mdx_batch_size_info"], minimum=1, maximum=64, value=1, step=1, interactive=True) 
-                        with gr.Row():
-                            kim_vocal_segments_size = gr.Slider(label=translations["segments_size"], info=translations["segments_size_info"], minimum=32, maximum=3072, value=256, step=32, interactive=True)
-                        with gr.Row():
-                            sample_rate0 = gr.Slider(minimum=8000, maximum=96000, step=1, value=44100, label=translations["sr"], info=translations["sr_info"], interactive=True)
-                with gr.Column():
-                    create_button = gr.Button(translations["createdataset"], variant="primary", scale=2, min_width=4000)
-                    with gr.Group():
-                        with gr.Row():
-                            clean_audio = gr.Checkbox(label=translations["clear_audio"], value=False, interactive=True)
-                            skip = gr.Checkbox(label=translations["skip"], value=False, interactive=True)
-                        with gr.Row():   
-                            dataset_clean_strength = gr.Slider(minimum=0, maximum=1, step=0.1, value=0.5, label=translations["clean_strength"], info=translations["clean_strength_info"], interactive=True, visible=clean_audio.value)
-                        with gr.Row():
-                            skip_start = gr.Textbox(label=translations["skip_start"], info=translations["skip_start_info"], value="", placeholder="0,...", interactive=True, visible=skip.value)
-                            skip_end = gr.Textbox(label=translations["skip_end"], info=translations["skip_end_info"], value="", placeholder="0,...", interactive=True, visible=skip.value)
-                    create_dataset_info = gr.Textbox(label=translations["create_dataset_info"], value="", interactive=False)
-            with gr.Row():
-                clean_audio.change(fn=visible, inputs=[clean_audio], outputs=[dataset_clean_strength])
-                skip.change(fn=lambda a: [valueEmpty_visible1(a)]*2, inputs=[skip], outputs=[skip_start, skip_end])
-            with gr.Row():
-                create_button.click(
-                    fn=create_dataset,
-                    inputs=[
-                        dataset_url, 
-                        output_dataset, 
-                        clean_audio, 
-                        dataset_clean_strength, 
-                        separator_reverb, 
-                        kim_vocal_version, 
-                        kim_vocal_overlap, 
-                        kim_vocal_segments_size, 
-                        denoise_mdx, 
-                        skip, 
-                        skip_start, 
-                        skip_end,
-                        kim_vocal_hop_length,
-                        kim_vocal_batch_size,
-                        sample_rate0
-                    ],
-                    outputs=[create_dataset_info],
-                    api_name="create_dataset"
-                )
-
-        with gr.Tab(label=translations["training_model"], visible=configs.get("training_tab", True)):
-            gr.Markdown(f"## {translations['training_model']}")
-            with gr.Row():
-                gr.Markdown(translations["training_markdown"])
-            with gr.Row():
-                with gr.Column():
-                    with gr.Row():
-                        with gr.Column():
-                            training_name = gr.Textbox(label=translations["modelname"], info=translations["training_model_name"], value="", placeholder=translations["modelname"], interactive=True)
-                            training_sr = gr.Radio(label=translations["sample_rate"], info=translations["sample_rate_info"], choices=["32k", "40k", "48k"], value="48k", interactive=True) 
-                            training_ver = gr.Radio(label=translations["training_version"], info=translations["training_version_info"], choices=["v1", "v2"], value="v2", interactive=True) 
-                            with gr.Row():
-                                clean_dataset = gr.Checkbox(label=translations["clear_dataset"], value=False, interactive=True)
-                                preprocess_cut = gr.Checkbox(label=translations["split_audio"], value=True, interactive=True)
-                                process_effects = gr.Checkbox(label=translations["preprocess_effect"], value=False, interactive=True)
-                                checkpointing1 = gr.Checkbox(label=translations["memory_efficient_training"], value=False, interactive=True)
-                                training_f0 = gr.Checkbox(label=translations["training_pitch"], value=True, interactive=True)
-                                upload = gr.Checkbox(label=translations["upload_dataset"], value=False, interactive=True)
-                            with gr.Row():
-                                clean_dataset_strength = gr.Slider(label=translations["clean_strength"], info=translations["clean_strength_info"], minimum=0, maximum=1, value=0.7, step=0.1, interactive=True, visible=clean_dataset.value)
-                        with gr.Column():
-                            preprocess_button = gr.Button(translations["preprocess_button"], scale=2)
-                            upload_dataset = gr.Files(label=translations["drop_audio"], file_types=[".wav", ".mp3", ".flac", ".ogg", ".opus", ".m4a", ".mp4", ".aac", ".alac", ".wma", ".aiff", ".webm", ".ac3"], visible=upload.value)
-                            preprocess_info = gr.Textbox(label=translations["preprocess_info"], value="", interactive=False)
-                with gr.Column():
-                    with gr.Row():
-                        with gr.Column():
-                            with gr.Accordion(label=translations["f0_method"], open=False):
-                                with gr.Group():
-                                    with gr.Row():
-                                        onnx_f0_mode2 = gr.Checkbox(label=translations["f0_onnx_mode"], info=translations["f0_onnx_mode_info"], value=False, interactive=True)
-                                        unlock_full_method4 = gr.Checkbox(label=translations["f0_unlock"], info=translations["f0_unlock_info"], value=False, interactive=True)
-                                    extract_method = gr.Radio(label=translations["f0_method"], info=translations["f0_method_info"], choices=method_f0, value="rmvpe", interactive=True)
-                                extract_hop_length = gr.Slider(label="Hop length", info=translations["hop_length_info"], minimum=1, maximum=512, value=128, step=1, interactive=True, visible=False)
-                            with gr.Accordion(label=translations["hubert_model"], open=False):
-                                with gr.Group():
-                                    embed_mode2 = gr.Radio(label=translations["embed_mode"], info=translations["embed_mode_info"], value="fairseq", choices=embedders_mode, interactive=True, visible=True)
-                                    extract_embedders = gr.Radio(label=translations["hubert_model"], info=translations["hubert_info"], choices=embedders_model, value="hubert_base", interactive=True)
-                                with gr.Row():
-                                    extract_embedders_custom = gr.Textbox(label=translations["modelname"], info=translations["modelname_info"], value="", placeholder="hubert_base", interactive=True, visible=extract_embedders.value == "custom")
-                        with gr.Column():
-                            extract_button = gr.Button(translations["extract_button"], scale=2)
-                            extract_info = gr.Textbox(label=translations["extract_info"], value="", interactive=False)
-                with gr.Column():
-                    with gr.Row():
-                        with gr.Column():
-                            total_epochs = gr.Slider(label=translations["total_epoch"], info=translations["total_epoch_info"], minimum=1, maximum=10000, value=300, step=1, interactive=True)
-                            save_epochs = gr.Slider(label=translations["save_epoch"], info=translations["save_epoch_info"], minimum=1, maximum=10000, value=50, step=1, interactive=True)
-                        with gr.Column():
-                            with gr.Row():
-                                index_button = gr.Button(f"3. {translations['create_index']}", variant="primary", scale=2)
-                                training_button = gr.Button(f"4. {translations['training_model']}", variant="primary", scale=2)
-                    with gr.Row():
-                        with gr.Accordion(label=translations["setting"], open=False):
-                            with gr.Row():
-                                index_algorithm = gr.Radio(label=translations["index_algorithm"], info=translations["index_algorithm_info"], choices=["Auto", "Faiss", "KMeans"], value="Auto", interactive=True)
-                            with gr.Row():
-                                custom_dataset = gr.Checkbox(label=translations["custom_dataset"], info=translations["custom_dataset_info"], value=False, interactive=True)
-                                overtraining_detector = gr.Checkbox(label=translations["overtraining_detector"], info=translations["overtraining_detector_info"], value=False, interactive=True)
-                                clean_up = gr.Checkbox(label=translations["cleanup_training"], info=translations["cleanup_training_info"], value=False, interactive=True)
-                                cache_in_gpu = gr.Checkbox(label=translations["cache_in_gpu"], info=translations["cache_in_gpu_info"], value=False, interactive=True)
-                            with gr.Column():
-                                dataset_path = gr.Textbox(label=translations["dataset_folder"], value="dataset", interactive=True, visible=custom_dataset.value)
-                            with gr.Column():
-                                threshold = gr.Slider(minimum=1, maximum=100, value=50, step=1, label=translations["threshold"], interactive=True, visible=overtraining_detector.value)
-                                with gr.Accordion(translations["setting_cpu_gpu"], open=False):
-                                    with gr.Column():
-                                        gpu_number = gr.Textbox(label=translations["gpu_number"], value=str("-".join(map(str, range(torch.cuda.device_count()))) if torch.cuda.is_available() else "-"), info=translations["gpu_number_info"], interactive=True)
-                                        gpu_info = gr.Textbox(label=translations["gpu_info"], value=get_gpu_info(), info=translations["gpu_info_2"], interactive=False)
-                                        cpu_core = gr.Slider(label=translations["cpu_core"], info=translations["cpu_core_info"], minimum=0, maximum=cpu_count(), value=cpu_count(), step=1, interactive=True)          
-                                        train_batch_size = gr.Slider(label=translations["batch_size"], info=translations["batch_size_info"], minimum=1, maximum=64, value=8, step=1, interactive=True)
-                            with gr.Row():
-                                save_only_latest = gr.Checkbox(label=translations["save_only_latest"], info=translations["save_only_latest_info"], value=True, interactive=True)
-                                save_every_weights = gr.Checkbox(label=translations["save_every_weights"], info=translations["save_every_weights_info"], value=True, interactive=True)
-                                not_use_pretrain = gr.Checkbox(label=translations["not_use_pretrain_2"], info=translations["not_use_pretrain_info"], value=False, interactive=True)
-                                custom_pretrain = gr.Checkbox(label=translations["custom_pretrain"], info=translations["custom_pretrain_info"], value=False, interactive=True)
-                            with gr.Row():
-                                vocoders = gr.Radio(label=translations["vocoder"], info=translations["vocoder_info"], choices=["Default", "MRF-HiFi-GAN", "RefineGAN"], value="Default", interactive=True) 
-                            with gr.Row():
-                                deterministic = gr.Checkbox(label=translations["deterministic"], info=translations["deterministic_info"], value=False, interactive=True)
-                                benchmark = gr.Checkbox(label=translations["benchmark"], info=translations["benchmark_info"], value=False, interactive=True)
-                            with gr.Row():
-                                model_author = gr.Textbox(label=translations["training_author"], info=translations["training_author_info"], value="", placeholder=translations["training_author"], interactive=True)
-                            with gr.Row():
-                                with gr.Column():
-                                    with gr.Accordion(translations["custom_pretrain_info"], open=False, visible=custom_pretrain.value and not not_use_pretrain.value) as pretrain_setting:
-                                        pretrained_D = gr.Dropdown(label=translations["pretrain_file"].format(dg="D"), choices=pretrainedD, value=pretrainedD[0] if len(pretrainedD) > 0 else '', interactive=True, allow_custom_value=True)
-                                        pretrained_G = gr.Dropdown(label=translations["pretrain_file"].format(dg="G"), choices=pretrainedG, value=pretrainedG[0] if len(pretrainedG) > 0 else '', interactive=True, allow_custom_value=True)
-                                        refesh_pretrain = gr.Button(translations["refesh"], scale=2)
-                    with gr.Row():
-                        training_info = gr.Textbox(label=translations["train_info"], value="", interactive=False)
-                    with gr.Row():
-                        with gr.Column():
-                            with gr.Accordion(translations["export_model"], open=False):
-                                with gr.Row():
-                                    model_file= gr.Dropdown(label=translations["model_name"], choices=model_name, value=model_name[0] if len(model_name) >= 1 else "", interactive=True, allow_custom_value=True)
-                                    index_file = gr.Dropdown(label=translations["index_path"], choices=index_path, value=index_path[0] if len(index_path) >= 1 else "", interactive=True, allow_custom_value=True)
-                                with gr.Row():
-                                    refesh_file = gr.Button(f"1. {translations['refesh']}", scale=2)
-                                    zip_model = gr.Button(translations["zip_model"], variant="primary", scale=2)
-                                with gr.Row():
-                                    zip_output = gr.File(label=translations["output_zip"], file_types=[".zip"], interactive=False, visible=False)
-            with gr.Row():
-                vocoders.change(fn=pitch_guidance_lock, inputs=[vocoders], outputs=[training_f0])
-                training_f0.change(fn=vocoders_lock, inputs=[training_f0, vocoders], outputs=[vocoders])
-                unlock_full_method4.change(fn=unlock_f0, inputs=[unlock_full_method4], outputs=[extract_method])
-            with gr.Row():
-                refesh_file.click(fn=change_models_choices, inputs=[], outputs=[model_file, index_file]) 
-                zip_model.click(fn=zip_file, inputs=[training_name, model_file, index_file], outputs=[zip_output])                
-                dataset_path.change(fn=lambda folder: os.makedirs(folder, exist_ok=True), inputs=[dataset_path], outputs=[])
-            with gr.Row():
-                upload.change(fn=visible, inputs=[upload], outputs=[upload_dataset]) 
-                overtraining_detector.change(fn=visible, inputs=[overtraining_detector], outputs=[threshold]) 
-                clean_dataset.change(fn=visible, inputs=[clean_dataset], outputs=[clean_dataset_strength])
-            with gr.Row():
-                custom_dataset.change(fn=lambda custom_dataset: [visible(custom_dataset), "dataset"],inputs=[custom_dataset], outputs=[dataset_path, dataset_path])
-                training_ver.change(fn=unlock_vocoder, inputs=[training_ver, vocoders], outputs=[vocoders])
-                vocoders.change(fn=unlock_ver, inputs=[training_ver, vocoders], outputs=[training_ver])
-                upload_dataset.upload(
-                    fn=lambda files, folder: [shutil.move(f.name, os.path.join(folder, os.path.split(f.name)[1])) for f in files] if folder != "" else gr_warning(translations["dataset_folder1"]),
-                    inputs=[upload_dataset, dataset_path], 
-                    outputs=[], 
-                    api_name="upload_dataset"
-                )           
-            with gr.Row():
-                not_use_pretrain.change(fn=lambda a, b: visible(a and not b), inputs=[custom_pretrain, not_use_pretrain], outputs=[pretrain_setting])
-                custom_pretrain.change(fn=lambda a, b: visible(a and not b), inputs=[custom_pretrain, not_use_pretrain], outputs=[pretrain_setting])
-                refesh_pretrain.click(fn=change_pretrained_choices, inputs=[], outputs=[pretrained_D, pretrained_G])
-            with gr.Row():
-                preprocess_button.click(
-                    fn=preprocess,
-                    inputs=[
-                        training_name, 
-                        training_sr, 
-                        cpu_core,
-                        preprocess_cut, 
-                        process_effects,
-                        dataset_path,
-                        clean_dataset,
-                        clean_dataset_strength
-                    ],
-                    outputs=[preprocess_info],
-                    api_name="preprocess"
-                )
-            with gr.Row():
-                embed_mode2.change(fn=visible_embedders, inputs=[embed_mode2], outputs=[extract_embedders])
-                extract_method.change(fn=hoplength_show, inputs=[extract_method], outputs=[extract_hop_length])
-                extract_embedders.change(fn=lambda extract_embedders: visible(extract_embedders == "custom"), inputs=[extract_embedders], outputs=[extract_embedders_custom])
-            with gr.Row():
-                extract_button.click(
-                    fn=extract,
-                    inputs=[
-                        training_name, 
-                        training_ver, 
-                        extract_method, 
-                        training_f0, 
-                        extract_hop_length, 
-                        cpu_core,
-                        gpu_number,
-                        training_sr, 
-                        extract_embedders, 
-                        extract_embedders_custom,
-                        onnx_f0_mode2,
-                        embed_mode2
-                    ],
-                    outputs=[extract_info],
-                    api_name="extract"
-                )
-            with gr.Row():
-                index_button.click(
-                    fn=create_index,
-                    inputs=[
-                        training_name, 
-                        training_ver, 
-                        index_algorithm
-                    ],
-                    outputs=[training_info],
-                    api_name="create_index"
-                )
-            with gr.Row():
-                training_button.click(
-                    fn=training,
-                    inputs=[
-                        training_name, 
-                        training_ver, 
-                        save_epochs, 
-                        save_only_latest, 
-                        save_every_weights, 
-                        total_epochs, 
-                        training_sr,
-                        train_batch_size, 
-                        gpu_number,
-                        training_f0,
-                        not_use_pretrain,
-                        custom_pretrain,
-                        pretrained_G,
-                        pretrained_D,
-                        overtraining_detector,
-                        threshold,
-                        clean_up,
-                        cache_in_gpu,
-                        model_author,
-                        vocoders,
-                        checkpointing1,
-                        deterministic, 
-                        benchmark
-                    ],
-                    outputs=[training_info],
-                    api_name="training_model"
-                )
-
-        with gr.Tab(label=translations["fushion"], visible=configs.get("fushion_tab", True)):
-            gr.Markdown(translations["fushion_markdown"])
-            with gr.Row():
-                gr.Markdown(translations["fushion_markdown_2"])
-            with gr.Row():
-                name_to_save = gr.Textbox(label=translations["modelname"], placeholder="Model.pth", value="", max_lines=1, interactive=True)
-            with gr.Row():
-                fushion_button = gr.Button(translations["fushion"], variant="primary", scale=4)
-            with gr.Column():
-                with gr.Row():
-                    model_a = gr.File(label=f"{translations['model_name']} 1", file_types=[".pth", ".onnx"]) 
-                    model_b = gr.File(label=f"{translations['model_name']} 2", file_types=[".pth", ".onnx"])
-                with gr.Row():
-                    model_path_a = gr.Textbox(label=f"{translations['model_path']} 1", value="", placeholder="assets/weights/Model_1.pth")
-                    model_path_b = gr.Textbox(label=f"{translations['model_path']} 2", value="", placeholder="assets/weights/Model_2.pth")
-            with gr.Row():
-                ratio = gr.Slider(minimum=0, maximum=1, label=translations["model_ratio"], info=translations["model_ratio_info"], value=0.5, interactive=True)
-            with gr.Row():
-                output_model = gr.File(label=translations["output_model_path"], file_types=[".pth", ".onnx"], interactive=False, visible=False)
-            with gr.Row():
-                model_a.upload(fn=lambda model: shutil.move(model.name, os.path.join("assets", "weights")), inputs=[model_a], outputs=[model_path_a])
-                model_b.upload(fn=lambda model: shutil.move(model.name, os.path.join("assets", "weights")), inputs=[model_b], outputs=[model_path_b])
-            with gr.Row():
-                fushion_button.click(
-                    fn=fushion_model,
-                    inputs=[
-                        name_to_save, 
-                        model_path_a, 
-                        model_path_b, 
-                        ratio
-                    ],
-                    outputs=[name_to_save, output_model],
-                    api_name="fushion_model"
-                )
-                fushion_button.click(fn=lambda: visible(True), inputs=[], outputs=[output_model])  
-
-        with gr.Tab(label=translations["read_model"], visible=configs.get("read_tab", True)):
-            gr.Markdown(translations["read_model_markdown"])
-            with gr.Row():
-                gr.Markdown(translations["read_model_markdown_2"])
-            with gr.Row():
-                model = gr.File(label=translations["drop_model"], file_types=[".pth", ".onnx"]) 
-            with gr.Row():
-                read_button = gr.Button(translations["readmodel"], variant="primary", scale=2)
-            with gr.Column():
-                model_path = gr.Textbox(label=translations["model_path"], value="", placeholder="assets/weights/Model.pth", info=translations["model_path_info"], interactive=True)
-                output_info = gr.Textbox(label=translations["modelinfo"], value="", interactive=False, scale=6)
-            with gr.Row():
-                model.upload(fn=lambda model: shutil.move(model.name, os.path.join("assets", "weights")), inputs=[model], outputs=[model_path])
-                read_button.click(
-                    fn=model_info,
-                    inputs=[model_path],
-                    outputs=[output_info],
-                    api_name="read_model"
-                )
-
-        with gr.Tab(label=translations["convert_model"], visible=configs.get("onnx_tab", True)):
-            gr.Markdown(translations["pytorch2onnx"])
-            with gr.Row():
-                gr.Markdown(translations["pytorch2onnx_markdown"])
-            with gr.Row():
-                model_pth_upload = gr.File(label=translations["drop_model"], file_types=[".pth"]) 
-            with gr.Row():
-                convert_onnx = gr.Button(translations["convert_model"], variant="primary", scale=2)
-            with gr.Row():
-                model_pth_path = gr.Textbox(label=translations["model_path"], value="", placeholder="assets/weights/Model.pth", info=translations["model_path_info"], interactive=True)
-            with gr.Row():
-                output_model2 = gr.File(label=translations["output_model_path"], file_types=[".pth", ".onnx"], interactive=False, visible=False)
-            with gr.Row():
-                model_pth_upload.upload(fn=lambda model_pth_upload: shutil.move(model_pth_upload.name, os.path.join("assets", "weights")), inputs=[model_pth_upload], outputs=[model_pth_path])
-                convert_onnx.click(
-                    fn=onnx_export,
-                    inputs=[model_pth_path],
-                    outputs=[output_model2, output_info],
-                    api_name="model_onnx_export"
-                )
-                convert_onnx.click(fn=lambda: visible(True), inputs=[], outputs=[output_model2])

main/app/tabs/utils/utils.py

@@ -1,305 +0,0 @@
-import gradio as gr
-from main.tools import huggingface
-from main.configs.config import Config
-from main.app.based.utils import *
-
-def utils_tabs():
-    with gr.TabItem("utils"):
-        with gr.Tabs():
-            with gr.TabItem(translations["audio_editing"], visible=False):
-                gr.Markdown(translations["audio_editing_info"])
-                with gr.Row():
-                    gr.Markdown(translations["audio_editing_markdown"])
-                with gr.Row():
-                    with gr.Column():
-                        with gr.Group():
-                            with gr.Row():
-                                save_compute = gr.Checkbox(label=translations["save_compute"], value=True, interactive=True)
-                            tar_prompt = gr.Textbox(label=translations["target_prompt"], info=translations["target_prompt_info"], placeholder="Piano and violin cover", lines=5, interactive=True)
-                    with gr.Column():
-                        cfg_scale_src = gr.Slider(value=3, minimum=0.5, maximum=25, label=translations["cfg_scale_src"], info=translations["cfg_scale_src_info"], interactive=True)
-                        cfg_scale_tar = gr.Slider(value=12, minimum=0.5, maximum=25, label=translations["cfg_scale_tar"], info=translations["cfg_scale_tar_info"], interactive=True)
-                with gr.Row():
-                    edit_button = gr.Button(translations["editing"], variant="primary")
-                with gr.Row():
-                    with gr.Column():
-                        drop_audio_file = gr.File(label=translations["drop_audio"], file_types=[".wav", ".mp3", ".flac", ".ogg", ".opus", ".m4a", ".mp4", ".aac", ".alac", ".wma", ".aiff", ".webm", ".ac3"])  
-                        display_audio = gr.Audio(show_download_button=True, interactive=False, label=translations["input_audio"])
-                    with gr.Column():
-                        with gr.Accordion(translations["input_output"], open=False):
-                            with gr.Column():
-                                export_audio_format = gr.Radio(label=translations["export_format"], info=translations["export_info"], choices=["wav", "mp3", "flac", "ogg", "opus", "m4a", "mp4", "aac", "alac", "wma", "aiff", "webm", "ac3"], value="wav", interactive=True)
-                                input_audiopath = gr.Dropdown(label=translations["audio_path"], value="", choices=paths_for_files, info=translations["provide_audio"], allow_custom_value=True, interactive=True)
-                                output_audiopath = gr.Textbox(label=translations["output_path"], value="audios/output.wav", placeholder="audios/output.wav", info=translations["output_path_info"], interactive=True)
-                            with gr.Column():
-                                refesh_audio = gr.Button(translations["refesh"])
-                        with gr.Accordion(translations["setting"], open=False):
-                            audioldm2_model = gr.Radio(label=translations["audioldm2_model"], info=translations["audioldm2_model_info"], choices=["audioldm2", "audioldm2-large", "audioldm2-music"], value="audioldm2-music", interactive=True)
-                            with gr.Row():
-                                src_prompt = gr.Textbox(label=translations["source_prompt"], lines=2, interactive=True, info=translations["source_prompt_info"], placeholder="A recording of a happy upbeat classical music piece")
-                            with gr.Row():
-                                with gr.Column(): 
-                                    audioldm2_sample_rate = gr.Slider(minimum=8000, maximum=96000, label=translations["sr"], info=translations["sr_info"], value=44100, step=1, interactive=True)
-                                    t_start = gr.Slider(minimum=15, maximum=85, value=45, step=1, label=translations["t_start"], interactive=True, info=translations["t_start_info"])
-                                    steps = gr.Slider(value=50, step=1, minimum=10, maximum=300, label=translations["steps_label"], info=translations["steps_info"], interactive=True)
-                with gr.Row():
-                    gr.Markdown(translations["output_audio"])
-                with gr.Row():
-                    output_audioldm2 = gr.Audio(show_download_button=True, interactive=False, label=translations["output_audio"])
-                with gr.Row():
-                    refesh_audio.click(fn=change_audios_choices, inputs=[input_audiopath], outputs=[input_audiopath])
-                    drop_audio_file.upload(fn=lambda audio_in: shutil.move(audio_in.name, os.path.join("audios")), inputs=[drop_audio_file], outputs=[input_audiopath])
-                    input_audiopath.change(fn=lambda audio: audio if os.path.isfile(audio) else None, inputs=[input_audiopath], outputs=[display_audio])
-                with gr.Row():
-                    edit_button.click(
-                        fn=run_audioldm2,
-                        inputs=[
-                            input_audiopath, 
-                            output_audiopath, 
-                            export_audio_format, 
-                            audioldm2_sample_rate, 
-                            audioldm2_model, 
-                            src_prompt, 
-                            tar_prompt, 
-                            steps, 
-                            cfg_scale_src, 
-                            cfg_scale_tar, 
-                            t_start, 
-                            save_compute
-                        ],
-                        outputs=[output_audioldm2],
-                        api_name="audioldm2"
-                    )
-
-            with gr.TabItem(translations["audio_effects"], visible=configs.get("effects_tab", True)):
-                gr.Markdown(translations["apply_audio_effects"])
-                with gr.Row():
-                    gr.Markdown(translations["audio_effects_edit"])
-                with gr.Row():
-                    with gr.Column():
-                        with gr.Row():
-                            reverb_check_box = gr.Checkbox(label=translations["reverb"], value=False, interactive=True)
-                            chorus_check_box = gr.Checkbox(label=translations["chorus"], value=False, interactive=True)
-                            delay_check_box = gr.Checkbox(label=translations["delay"], value=False, interactive=True)
-                            phaser_check_box = gr.Checkbox(label=translations["phaser"], value=False, interactive=True)
-                            compressor_check_box = gr.Checkbox(label=translations["compressor"], value=False, interactive=True)
-                            more_options = gr.Checkbox(label=translations["more_option"], value=False, interactive=True)    
-                with gr.Row():
-                    with gr.Accordion(translations["input_output"], open=False):
-                        with gr.Row():
-                            upload_audio = gr.File(label=translations["drop_audio"], file_types=[".wav", ".mp3", ".flac", ".ogg", ".opus", ".m4a", ".mp4", ".aac", ".alac", ".wma", ".aiff", ".webm", ".ac3"])
-                        with gr.Row():
-                            audio_in_path = gr.Dropdown(label=translations["input_audio"], value="", choices=paths_for_files, info=translations["provide_audio"], interactive=True, allow_custom_value=True)
-                            audio_out_path = gr.Textbox(label=translations["output_audio"], value="audios/audio_effects.wav", placeholder="audios/audio_effects.wav", info=translations["provide_output"], interactive=True)
-                        with gr.Row():
-                            with gr.Column():
-                                audio_combination = gr.Checkbox(label=translations["merge_instruments"], value=False, interactive=True)
-                                audio_combination_input = gr.Dropdown(label=translations["input_audio"], value="", choices=paths_for_files, info=translations["provide_audio"], interactive=True, allow_custom_value=True, visible=audio_combination.value)
-                        with gr.Row():
-                            audio_effects_refesh = gr.Button(translations["refesh"])
-                        with gr.Row():
-                            audio_output_format = gr.Radio(label=translations["export_format"], info=translations["export_info"], choices=["wav", "mp3", "flac", "ogg", "opus", "m4a", "mp4", "aac", "alac", "wma", "aiff", "webm", "ac3"], value="wav", interactive=True)
-                with gr.Row():
-                    apply_effects_button = gr.Button(translations["apply"], variant="primary", scale=2)
-                with gr.Row():
-                    with gr.Column():
-                        with gr.Row():
-                            with gr.Accordion(translations["reverb"], open=False, visible=reverb_check_box.value) as reverb_accordion:
-                                reverb_freeze_mode = gr.Checkbox(label=translations["reverb_freeze"], info=translations["reverb_freeze_info"], value=False, interactive=True)
-                                reverb_room_size = gr.Slider(minimum=0, maximum=1, step=0.01, value=0.15, label=translations["room_size"], info=translations["room_size_info"], interactive=True)
-                                reverb_damping = gr.Slider(minimum=0, maximum=1, step=0.01, value=0.7, label=translations["damping"], info=translations["damping_info"], interactive=True)
-                                reverb_wet_level = gr.Slider(minimum=0, maximum=1, step=0.01, value=0.2, label=translations["wet_level"], info=translations["wet_level_info"], interactive=True)
-                                reverb_dry_level = gr.Slider(minimum=0, maximum=1, step=0.01, value=0.8, label=translations["dry_level"], info=translations["dry_level_info"], interactive=True)
-                                reverb_width = gr.Slider(minimum=0, maximum=1, step=0.01, value=1, label=translations["width"], info=translations["width_info"], interactive=True)
-                        with gr.Row():
-                            with gr.Accordion(translations["chorus"], open=False, visible=chorus_check_box.value) as chorus_accordion:
-                                chorus_depth = gr.Slider(minimum=0, maximum=1, step=0.01, value=0.5, label=translations["chorus_depth"], info=translations["chorus_depth_info"], interactive=True)
-                                chorus_rate_hz = gr.Slider(minimum=0.1, maximum=10, step=0.1, value=1.5, label=translations["chorus_rate_hz"], info=translations["chorus_rate_hz_info"], interactive=True)
-                                chorus_mix = gr.Slider(minimum=0, maximum=1, step=0.01, value=0.5, label=translations["chorus_mix"], info=translations["chorus_mix_info"], interactive=True)
-                                chorus_centre_delay_ms = gr.Slider(minimum=0, maximum=50, step=1, value=10, label=translations["chorus_centre_delay_ms"], info=translations["chorus_centre_delay_ms_info"], interactive=True)
-                                chorus_feedback = gr.Slider(minimum=-1, maximum=1, step=0.01, value=0, label=translations["chorus_feedback"], info=translations["chorus_feedback_info"], interactive=True)
-                        with gr.Row():
-                            with gr.Accordion(translations["delay"], open=False, visible=delay_check_box.value) as delay_accordion:
-                                delay_second = gr.Slider(minimum=0, maximum=5, step=0.01, value=0.5, label=translations["delay_seconds"], info=translations["delay_seconds_info"], interactive=True)
-                                delay_feedback = gr.Slider(minimum=0, maximum=1, step=0.01, value=0.5, label=translations["delay_feedback"], info=translations["delay_feedback_info"], interactive=True)
-                                delay_mix = gr.Slider(minimum=0, maximum=1, step=0.01, value=0.5, label=translations["delay_mix"], info=translations["delay_mix_info"], interactive=True)
-                    with gr.Column():
-                        with gr.Row():
-                            with gr.Accordion(translations["more_option"], open=False, visible=more_options.value) as more_accordion:
-                                with gr.Row():
-                                    fade = gr.Checkbox(label=translations["fade"], value=False, interactive=True)
-                                    bass_or_treble = gr.Checkbox(label=translations["bass_or_treble"], value=False, interactive=True)
-                                    limiter = gr.Checkbox(label=translations["limiter"], value=False, interactive=True)
-                                    resample_checkbox = gr.Checkbox(label=translations["resample"], value=False, interactive=True)
-                                with gr.Row():
-                                    distortion_checkbox = gr.Checkbox(label=translations["distortion"], value=False, interactive=True)
-                                    gain_checkbox = gr.Checkbox(label=translations["gain"], value=False, interactive=True)
-                                    bitcrush_checkbox = gr.Checkbox(label=translations["bitcrush"], value=False, interactive=True)
-                                    clipping_checkbox = gr.Checkbox(label=translations["clipping"], value=False, interactive=True)
-                                with gr.Accordion(translations["fade"], open=True, visible=fade.value) as fade_accordion:
-                                    with gr.Row():
-                                        fade_in = gr.Slider(minimum=0, maximum=10000, step=100, value=0, label=translations["fade_in"], info=translations["fade_in_info"], interactive=True)
-                                        fade_out = gr.Slider(minimum=0, maximum=10000, step=100, value=0, label=translations["fade_out"], info=translations["fade_out_info"], interactive=True)
-                                with gr.Accordion(translations["bass_or_treble"], open=True, visible=bass_or_treble.value) as bass_treble_accordion:
-                                    with gr.Row():
-                                        bass_boost = gr.Slider(minimum=0, maximum=20, step=1, value=0, label=translations["bass_boost"], info=translations["bass_boost_info"], interactive=True)
-                                        bass_frequency = gr.Slider(minimum=20, maximum=200, step=10, value=100, label=translations["bass_frequency"], info=translations["bass_frequency_info"], interactive=True)
-                                    with gr.Row():
-                                        treble_boost = gr.Slider(minimum=0, maximum=20, step=1, value=0, label=translations["treble_boost"], info=translations["treble_boost_info"], interactive=True)
-                                        treble_frequency = gr.Slider(minimum=1000, maximum=10000, step=500, value=3000, label=translations["treble_frequency"], info=translations["treble_frequency_info"], interactive=True)
-                                with gr.Accordion(translations["limiter"], open=True, visible=limiter.value) as limiter_accordion:
-                                    with gr.Row():
-                                        limiter_threashold_db = gr.Slider(minimum=-60, maximum=0, step=1, value=-1, label=translations["limiter_threashold_db"], info=translations["limiter_threashold_db_info"], interactive=True)
-                                        limiter_release_ms = gr.Slider(minimum=10, maximum=1000, step=1, value=100, label=translations["limiter_release_ms"], info=translations["limiter_release_ms_info"], interactive=True)
-                                with gr.Column():
-                                    pitch_shift_semitones = gr.Slider(minimum=-20, maximum=20, step=1, value=0, label=translations["pitch"], info=translations["pitch_info"], interactive=True)
-                                    audio_effect_resample_sr = gr.Slider(minimum=0, maximum=96000, step=1, value=0, label=translations["resample"], info=translations["resample_info"], interactive=True, visible=resample_checkbox.value)
-                                    distortion_drive_db = gr.Slider(minimum=0, maximum=50, step=1, value=20, label=translations["distortion"], info=translations["distortion_info"], interactive=True, visible=distortion_checkbox.value)
-                                    gain_db = gr.Slider(minimum=-60, maximum=60, step=1, value=0, label=translations["gain"], info=translations["gain_info"], interactive=True, visible=gain_checkbox.value)
-                                    clipping_threashold_db = gr.Slider(minimum=-60, maximum=0, step=1, value=-1, label=translations["clipping_threashold_db"], info=translations["clipping_threashold_db_info"], interactive=True, visible=clipping_checkbox.value)
-                                    bitcrush_bit_depth = gr.Slider(minimum=1, maximum=24, step=1, value=16, label=translations["bitcrush_bit_depth"], info=translations["bitcrush_bit_depth_info"], interactive=True, visible=bitcrush_checkbox.value)
-                        with gr.Row():
-                            with gr.Accordion(translations["phaser"], open=False, visible=phaser_check_box.value) as phaser_accordion:
-                                phaser_depth = gr.Slider(minimum=0, maximum=1, step=0.01, value=0.5, label=translations["phaser_depth"], info=translations["phaser_depth_info"], interactive=True)
-                                phaser_rate_hz = gr.Slider(minimum=0.1, maximum=10, step=0.1, value=1, label=translations["phaser_rate_hz"], info=translations["phaser_rate_hz_info"], interactive=True)
-                                phaser_mix = gr.Slider(minimum=0, maximum=1, step=0.01, value=0.5, label=translations["phaser_mix"], info=translations["phaser_mix_info"], interactive=True)
-                                phaser_centre_frequency_hz = gr.Slider(minimum=50, maximum=5000, step=10, value=1000, label=translations["phaser_centre_frequency_hz"], info=translations["phaser_centre_frequency_hz_info"], interactive=True)
-                                phaser_feedback = gr.Slider(minimum=-1, maximum=1, step=0.01, value=0, label=translations["phaser_feedback"], info=translations["phaser_feedback_info"], interactive=True)
-                        with gr.Row():
-                            with gr.Accordion(translations["compressor"], open=False, visible=compressor_check_box.value) as compressor_accordion:
-                                compressor_threashold_db = gr.Slider(minimum=-60, maximum=0, step=1, value=-20, label=translations["compressor_threashold_db"], info=translations["compressor_threashold_db_info"], interactive=True)
-                                compressor_ratio = gr.Slider(minimum=1, maximum=20, step=0.1, value=1, label=translations["compressor_ratio"], info=translations["compressor_ratio_info"], interactive=True)
-                                compressor_attack_ms = gr.Slider(minimum=0.1, maximum=100, step=0.1, value=10, label=translations["compressor_attack_ms"], info=translations["compressor_attack_ms_info"], interactive=True)
-                                compressor_release_ms = gr.Slider(minimum=10, maximum=1000, step=1, value=100, label=translations["compressor_release_ms"], info=translations["compressor_release_ms_info"], interactive=True)   
-                with gr.Row():
-                    gr.Markdown(translations["output_audio"])
-                with gr.Row():
-                    audio_play_input = gr.Audio(show_download_button=True, interactive=False, label=translations["input_audio"])
-                    audio_play_output = gr.Audio(show_download_button=True, interactive=False, label=translations["output_audio"])
-                with gr.Row():
-                    reverb_check_box.change(fn=visible, inputs=[reverb_check_box], outputs=[reverb_accordion])
-                    chorus_check_box.change(fn=visible, inputs=[chorus_check_box], outputs=[chorus_accordion])
-                    delay_check_box.change(fn=visible, inputs=[delay_check_box], outputs=[delay_accordion])
-                with gr.Row():
-                    compressor_check_box.change(fn=visible, inputs=[compressor_check_box], outputs=[compressor_accordion])
-                    phaser_check_box.change(fn=visible, inputs=[phaser_check_box], outputs=[phaser_accordion])
-                    more_options.change(fn=visible, inputs=[more_options], outputs=[more_accordion])
-                with gr.Row():
-                    fade.change(fn=visible, inputs=[fade], outputs=[fade_accordion])
-                    bass_or_treble.change(fn=visible, inputs=[bass_or_treble], outputs=[bass_treble_accordion])
-                    limiter.change(fn=visible, inputs=[limiter], outputs=[limiter_accordion])
-                    resample_checkbox.change(fn=visible, inputs=[resample_checkbox], outputs=[audio_effect_resample_sr])
-                with gr.Row():
-                    distortion_checkbox.change(fn=visible, inputs=[distortion_checkbox], outputs=[distortion_drive_db])
-                    gain_checkbox.change(fn=visible, inputs=[gain_checkbox], outputs=[gain_db])
-                    clipping_checkbox.change(fn=visible, inputs=[clipping_checkbox], outputs=[clipping_threashold_db])
-                    bitcrush_checkbox.change(fn=visible, inputs=[bitcrush_checkbox], outputs=[bitcrush_bit_depth])
-                with gr.Row():
-                    upload_audio.upload(fn=lambda audio_in: shutil.move(audio_in.name, os.path.join("audios")), inputs=[upload_audio], outputs=[audio_in_path])
-                    audio_in_path.change(fn=lambda audio: audio if audio else None, inputs=[audio_in_path], outputs=[audio_play_input])
-                    audio_effects_refesh.click(fn=lambda a, b: [change_audios_choices(a), change_audios_choices(b)], inputs=[audio_in_path, audio_combination_input], outputs=[audio_in_path, audio_combination_input])
-                with gr.Row():
-                    more_options.change(fn=lambda: [False]*8, inputs=[], outputs=[fade, bass_or_treble, limiter, resample_checkbox, distortion_checkbox, gain_checkbox, clipping_checkbox, bitcrush_checkbox])
-                    audio_combination.change(fn=visible, inputs=[audio_combination], outputs=[audio_combination_input])
-                with gr.Row():
-                    apply_effects_button.click(
-                        fn=audio_effects,
-                        inputs=[
-                            audio_in_path, 
-                            audio_out_path, 
-                            resample_checkbox, 
-                            audio_effect_resample_sr, 
-                            chorus_depth, 
-                            chorus_rate_hz, 
-                            chorus_mix, 
-                            chorus_centre_delay_ms, 
-                            chorus_feedback, 
-                            distortion_drive_db, 
-                            reverb_room_size, 
-                            reverb_damping, 
-                            reverb_wet_level, 
-                            reverb_dry_level, 
-                            reverb_width, 
-                            reverb_freeze_mode, 
-                            pitch_shift_semitones, 
-                            delay_second, 
-                            delay_feedback, 
-                            delay_mix, 
-                            compressor_threashold_db, 
-                            compressor_ratio, 
-                            compressor_attack_ms, 
-                            compressor_release_ms, 
-                            limiter_threashold_db, 
-                            limiter_release_ms, 
-                            gain_db, 
-                            bitcrush_bit_depth, 
-                            clipping_threashold_db, 
-                            phaser_rate_hz, 
-                            phaser_depth, 
-                            phaser_centre_frequency_hz, 
-                            phaser_feedback, 
-                            phaser_mix, 
-                            bass_boost, 
-                            bass_frequency, 
-                            treble_boost, 
-                            treble_frequency, 
-                            fade_in, 
-                            fade_out, 
-                            audio_output_format, 
-                            chorus_check_box, 
-                            distortion_checkbox, 
-                            reverb_check_box, 
-                            delay_check_box, 
-                            compressor_check_box, 
-                            limiter, 
-                            gain_checkbox, 
-                            bitcrush_checkbox, 
-                            clipping_checkbox, 
-                            phaser_check_box, 
-                            bass_or_treble, 
-                            fade,
-                            audio_combination,
-                            audio_combination_input
-                        ],
-                        outputs=[audio_play_output],
-                        api_name="audio_effects"
-                    )
-
-            with gr.TabItem(translations["f0_extractor_tab"], visible=configs.get("f0_extractor_tab", True)):
-                gr.Markdown(translations["f0_extractor_markdown"])
-                with gr.Row():
-                    gr.Markdown(translations["f0_extractor_markdown_2"])
-                with gr.Row():
-                    extractor_button = gr.Button(translations["extract_button"].replace("2. ", ""), variant="primary")
-                with gr.Row():
-                    with gr.Column():
-                        upload_audio_file = gr.File(label=translations["drop_audio"], file_types=[".wav", ".mp3", ".flac", ".ogg", ".opus", ".m4a", ".mp4", ".aac", ".alac", ".wma", ".aiff", ".webm", ".ac3"])
-                        audioplay = gr.Audio(show_download_button=True, interactive=False, label=translations["input_audio"])
-                    with gr.Column():
-                        with gr.Accordion(translations["f0_method"], open=False):
-                            with gr.Group():
-                                onnx_f0_mode3 = gr.Checkbox(label=translations["f0_onnx_mode"], info=translations["f0_onnx_mode_info"], value=False, interactive=True)
-                                f0_method_extract = gr.Radio(label=translations["f0_method"], info=translations["f0_method_info"], choices=method_f0, value="rmvpe", interactive=True)
-                        with gr.Accordion(translations["audio_path"], open=True):
-                            input_audio_path = gr.Dropdown(label=translations["audio_path"], value="", choices=paths_for_files, allow_custom_value=True, interactive=True)
-                            refesh_audio_button = gr.Button(translations["refesh"])
-                with gr.Row():
-                    gr.Markdown("___")
-                with gr.Row():
-                    file_output = gr.File(label="", file_types=[".txt"], interactive=False)
-                    image_output = gr.Image(label="", interactive=False, show_download_button=True)
-                with gr.Row():
-                    upload_audio_file.upload(fn=lambda audio_in: shutil.move(audio_in.name, os.path.join("audios")), inputs=[upload_audio_file], outputs=[input_audio_path])
-                    input_audio_path.change(fn=lambda audio: audio if os.path.isfile(audio) else None, inputs=[input_audio_path], outputs=[audioplay])
-                    refesh_audio_button.click(fn=change_audios_choices, inputs=[input_audio_path], outputs=[input_audio_path])
-                with gr.Row():
-                    extractor_button.click(
-                        fn=f0_extract,
-                        inputs=[
-                            input_audio_path,
-                            f0_method_extract,
-                            onnx_f0_mode3
-                        ],
-                        outputs=[file_output, image_output],
-                        api_name="f0_extract"
-                    )

main/app/tensorboard.py

@@ -1,30 +0,0 @@
-import os
-import sys
-import json
-import logging
-import webbrowser
-
-from tensorboard import program
-
-sys.path.append(os.getcwd())
-
-from main.configs.config import Config
-translations = Config().translations
-
-with open(os.path.join("main", "configs", "config.json"), "r") as f:
-    configs = json.load(f)
-
-def launch_tensorboard():
-    for l in ["root", "tensorboard"]:
-        logging.getLogger(l).setLevel(logging.ERROR)
-
-    tb = program.TensorBoard()
-    tb.configure(argv=[None, "--logdir", "assets/logs", f"--port={configs['tensorboard_port']}"])
-    url = tb.launch()
-
-    print(f"{translations['tensorboard_url']}: {url}")
-    if "--open" in sys.argv: webbrowser.open(url)
-
-    return f"{translations['tensorboard_url']}: {url}"
-
-if __name__ == "__main__": launch_tensorboard()

main/configs/config.json

@@ -1,549 +0,0 @@
-{
-    "language": "en-US",
-    "support_language": [
-        "en-US",
-        "id_Id",
-        "ja-JP",
-        "vi-VN"
-    ],
-    "theme": "gradio/default",
-    "themes": [
-        "NoCrypt/miku",
-        "gstaff/xkcd",
-        "JohnSmith9982/small_and_pretty",
-        "ParityError/Interstellar",
-        "earneleh/paris",
-        "shivi/calm_seafoam",
-        "Hev832/Applio",
-        "YTheme/Minecraft",
-        "gstaff/sketch",
-        "SebastianBravo/simci_css",
-        "allenai/gradio-theme",
-        "Nymbo/Nymbo_Theme_5",
-        "lone17/kotaemon",
-        "Zarkel/IBM_Carbon_Theme",
-        "SherlockRamos/Feliz",
-        "freddyaboulton/dracula_revamped",
-        "freddyaboulton/bad-theme-space",
-        "gradio/dracula_revamped",
-        "abidlabs/dracula_revamped",
-        "gradio/dracula_test",
-        "gradio/seafoam",
-        "gradio/glass",
-        "gradio/monochrome",
-        "gradio/soft",
-        "gradio/default",
-        "gradio/base",
-        "abidlabs/pakistan",
-        "dawood/microsoft_windows",
-        "ysharma/steampunk",
-        "ysharma/huggingface",
-        "abidlabs/Lime",
-        "freddyaboulton/this-theme-does-not-exist-2",
-        "aliabid94/new-theme",
-        "aliabid94/test2",
-        "aliabid94/test3",
-        "aliabid94/test4",
-        "abidlabs/banana",
-        "freddyaboulton/test-blue",
-        "gstaff/whiteboard",
-        "ysharma/llamas",
-        "abidlabs/font-test",
-        "YenLai/Superhuman",
-        "bethecloud/storj_theme",
-        "sudeepshouche/minimalist",
-        "knotdgaf/gradiotest",
-        "ParityError/Anime",
-        "Ajaxon6255/Emerald_Isle",
-        "ParityError/LimeFace",
-        "finlaymacklon/smooth_slate",
-        "finlaymacklon/boxy_violet",
-        "derekzen/stardust",
-        "EveryPizza/Cartoony-Gradio-Theme",
-        "Ifeanyi/Cyanister",
-        "Tshackelton/IBMPlex-DenseReadable",
-        "snehilsanyal/scikit-learn",
-        "Himhimhim/xkcd",
-        "nota-ai/theme",
-        "rawrsor1/Everforest",
-        "rottenlittlecreature/Moon_Goblin",
-        "abidlabs/test-yellow",
-        "abidlabs/test-yellow3",
-        "idspicQstitho/dracula_revamped",
-        "kfahn/AnimalPose",
-        "HaleyCH/HaleyCH_Theme",
-        "simulKitke/dracula_test",
-        "braintacles/CrimsonNight",
-        "wentaohe/whiteboardv2",
-        "reilnuud/polite",
-        "remilia/Ghostly",
-        "Franklisi/darkmode",
-        "coding-alt/soft",
-        "xiaobaiyuan/theme_land",
-        "step-3-profit/Midnight-Deep",
-        "xiaobaiyuan/theme_demo",
-        "Taithrah/Minimal",
-        "Insuz/SimpleIndigo",
-        "zkunn/Alipay_Gradio_theme",
-        "Insuz/Mocha",
-        "xiaobaiyuan/theme_brief",
-        "Ama434/434-base-Barlow",
-        "Ama434/def_barlow",
-        "Ama434/neutral-barlow",
-        "dawood/dracula_test",
-        "nuttea/Softblue",
-        "BlueDancer/Alien_Diffusion",
-        "naughtondale/monochrome",
-        "Dagfinn1962/standard",
-        "default"
-    ],
-    "mdx_model": [
-        "Main_340",
-        "Main_390",
-        "Main_406",
-        "Main_427",
-        "Main_438",
-        "Inst_full_292",
-        "Inst_HQ_1",
-        "Inst_HQ_2",
-        "Inst_HQ_3",
-        "Inst_HQ_4",
-        "Inst_HQ_5",
-        "Kim_Vocal_1",
-        "Kim_Vocal_2",
-        "Kim_Inst",
-        "Inst_187_beta",
-        "Inst_82_beta",
-        "Inst_90_beta",
-        "Voc_FT",
-        "Crowd_HQ",
-        "Inst_1",
-        "Inst_2",
-        "Inst_3",
-        "MDXNET_1_9703",
-        "MDXNET_2_9682",
-        "MDXNET_3_9662",
-        "Inst_Main",
-        "MDXNET_Main",
-        "MDXNET_9482"
-    ],
-    "demucs_model": [
-        "HT-Normal",
-        "HT-Tuned",
-        "HD_MMI",
-        "HT_6S"
-    ],
-    "edge_tts": [
-        "af-ZA-AdriNeural",
-        "af-ZA-WillemNeural",
-        "sq-AL-AnilaNeural",
-        "sq-AL-IlirNeural",
-        "am-ET-AmehaNeural",
-        "am-ET-MekdesNeural",
-        "ar-DZ-AminaNeural",
-        "ar-DZ-IsmaelNeural",
-        "ar-BH-AliNeural",
-        "ar-BH-LailaNeural",
-        "ar-EG-SalmaNeural",
-        "ar-EG-ShakirNeural",
-        "ar-IQ-BasselNeural",
-        "ar-IQ-RanaNeural",
-        "ar-JO-SanaNeural",
-        "ar-JO-TaimNeural",
-        "ar-KW-FahedNeural",
-        "ar-KW-NouraNeural",
-        "ar-LB-LaylaNeural",
-        "ar-LB-RamiNeural",
-        "ar-LY-ImanNeural",
-        "ar-LY-OmarNeural",
-        "ar-MA-JamalNeural",
-        "ar-MA-MounaNeural",
-        "ar-OM-AbdullahNeural",
-        "ar-OM-AyshaNeural",
-        "ar-QA-AmalNeural",
-        "ar-QA-MoazNeural",
-        "ar-SA-HamedNeural",
-        "ar-SA-ZariyahNeural",
-        "ar-SY-AmanyNeural",
-        "ar-SY-LaithNeural",
-        "ar-TN-HediNeural",
-        "ar-TN-ReemNeural",
-        "ar-AE-FatimaNeural",
-        "ar-AE-HamdanNeural",
-        "ar-YE-MaryamNeural",
-        "ar-YE-SalehNeural",
-        "az-AZ-BabekNeural",
-        "az-AZ-BanuNeural",
-        "bn-BD-NabanitaNeural",
-        "bn-BD-PradeepNeural",
-        "bn-IN-BashkarNeural",
-        "bn-IN-TanishaaNeural",
-        "bs-BA-GoranNeural",
-        "bs-BA-VesnaNeural",
-        "bg-BG-BorislavNeural",
-        "bg-BG-KalinaNeural",
-        "my-MM-NilarNeural",
-        "my-MM-ThihaNeural",
-        "ca-ES-EnricNeural",
-        "ca-ES-JoanaNeural",
-        "zh-HK-HiuGaaiNeural",
-        "zh-HK-HiuMaanNeural",
-        "zh-HK-WanLungNeural",
-        "zh-CN-XiaoxiaoNeural",
-        "zh-CN-XiaoyiNeural",
-        "zh-CN-YunjianNeural",
-        "zh-CN-YunxiNeural",
-        "zh-CN-YunxiaNeural",
-        "zh-CN-YunyangNeural",
-        "zh-CN-liaoning-XiaobeiNeural",
-        "zh-TW-HsiaoChenNeural",
-        "zh-TW-YunJheNeural",
-        "zh-TW-HsiaoYuNeural",
-        "zh-CN-shaanxi-XiaoniNeural",
-        "hr-HR-GabrijelaNeural",
-        "hr-HR-SreckoNeural",
-        "cs-CZ-AntoninNeural",
-        "cs-CZ-VlastaNeural",
-        "da-DK-ChristelNeural",
-        "da-DK-JeppeNeural",
-        "nl-BE-ArnaudNeural",
-        "nl-BE-DenaNeural",
-        "nl-NL-ColetteNeural",
-        "nl-NL-FennaNeural",
-        "nl-NL-MaartenNeural",
-        "en-AU-NatashaNeural",
-        "en-AU-WilliamNeural",
-        "en-CA-ClaraNeural",
-        "en-CA-LiamNeural",
-        "en-HK-SamNeural",
-        "en-HK-YanNeural",
-        "en-IN-NeerjaExpressiveNeural",
-        "en-IN-NeerjaNeural",
-        "en-IN-PrabhatNeural",
-        "en-IE-ConnorNeural",
-        "en-IE-EmilyNeural",
-        "en-KE-AsiliaNeural",
-        "en-KE-ChilembaNeural",
-        "en-NZ-MitchellNeural",
-        "en-NZ-MollyNeural",
-        "en-NG-AbeoNeural",
-        "en-NG-EzinneNeural",
-        "en-PH-JamesNeural",
-        "en-PH-RosaNeural",
-        "en-SG-LunaNeural",
-        "en-SG-WayneNeural",
-        "en-ZA-LeahNeural",
-        "en-ZA-LukeNeural",
-        "en-TZ-ElimuNeural",
-        "en-TZ-ImaniNeural",
-        "en-GB-LibbyNeural",
-        "en-GB-MaisieNeural",
-        "en-GB-RyanNeural",
-        "en-GB-SoniaNeural",
-        "en-GB-ThomasNeural",
-        "en-US-AvaMultilingualNeural",
-        "en-US-AndrewMultilingualNeural",
-        "en-US-EmmaMultilingualNeural",
-        "en-US-BrianMultilingualNeural",
-        "en-US-AvaNeural",
-        "en-US-AndrewNeural",
-        "en-US-EmmaNeural",
-        "en-US-BrianNeural",
-        "en-US-AnaNeural",
-        "en-US-AriaNeural",
-        "en-US-ChristopherNeural",
-        "en-US-EricNeural",
-        "en-US-GuyNeural",
-        "en-US-JennyNeural",
-        "en-US-MichelleNeural",
-        "en-US-RogerNeural",
-        "en-US-SteffanNeural",
-        "et-EE-AnuNeural",
-        "et-EE-KertNeural",
-        "fil-PH-AngeloNeural",
-        "fil-PH-BlessicaNeural",
-        "fi-FI-HarriNeural",
-        "fi-FI-NooraNeural",
-        "fr-BE-CharlineNeural",
-        "fr-BE-GerardNeural",
-        "fr-CA-ThierryNeural",
-        "fr-CA-AntoineNeural",
-        "fr-CA-JeanNeural",
-        "fr-CA-SylvieNeural",
-        "fr-FR-VivienneMultilingualNeural",
-        "fr-FR-RemyMultilingualNeural",
-        "fr-FR-DeniseNeural",
-        "fr-FR-EloiseNeural",
-        "fr-FR-HenriNeural",
-        "fr-CH-ArianeNeural",
-        "fr-CH-FabriceNeural",
-        "gl-ES-RoiNeural",
-        "gl-ES-SabelaNeural",
-        "ka-GE-EkaNeural",
-        "ka-GE-GiorgiNeural",
-        "de-AT-IngridNeural",
-        "de-AT-JonasNeural",
-        "de-DE-SeraphinaMultilingualNeural",
-        "de-DE-FlorianMultilingualNeural",
-        "de-DE-AmalaNeural",
-        "de-DE-ConradNeural",
-        "de-DE-KatjaNeural",
-        "de-DE-KillianNeural",
-        "de-CH-JanNeural",
-        "de-CH-LeniNeural",
-        "el-GR-AthinaNeural",
-        "el-GR-NestorasNeural",
-        "gu-IN-DhwaniNeural",
-        "gu-IN-NiranjanNeural",
-        "he-IL-AvriNeural",
-        "he-IL-HilaNeural",
-        "hi-IN-MadhurNeural",
-        "hi-IN-SwaraNeural",
-        "hu-HU-NoemiNeural",
-        "hu-HU-TamasNeural",
-        "is-IS-GudrunNeural",
-        "is-IS-GunnarNeural",
-        "id-ID-ArdiNeural",
-        "id-ID-GadisNeural",
-        "ga-IE-ColmNeural",
-        "ga-IE-OrlaNeural",
-        "it-IT-GiuseppeNeural",
-        "it-IT-DiegoNeural",
-        "it-IT-ElsaNeural",
-        "it-IT-IsabellaNeural",
-        "ja-JP-KeitaNeural",
-        "ja-JP-NanamiNeural",
-        "jv-ID-DimasNeural",
-        "jv-ID-SitiNeural",
-        "kn-IN-GaganNeural",
-        "kn-IN-SapnaNeural",
-        "kk-KZ-AigulNeural",
-        "kk-KZ-DauletNeural",
-        "km-KH-PisethNeural",
-        "km-KH-SreymomNeural",
-        "ko-KR-HyunsuNeural",
-        "ko-KR-InJoonNeural",
-        "ko-KR-SunHiNeural",
-        "lo-LA-ChanthavongNeural",
-        "lo-LA-KeomanyNeural",
-        "lv-LV-EveritaNeural",
-        "lv-LV-NilsNeural",
-        "lt-LT-LeonasNeural",
-        "lt-LT-OnaNeural",
-        "mk-MK-AleksandarNeural",
-        "mk-MK-MarijaNeural",
-        "ms-MY-OsmanNeural",
-        "ms-MY-YasminNeural",
-        "ml-IN-MidhunNeural",
-        "ml-IN-SobhanaNeural",
-        "mt-MT-GraceNeural",
-        "mt-MT-JosephNeural",
-        "mr-IN-AarohiNeural",
-        "mr-IN-ManoharNeural",
-        "mn-MN-BataaNeural",
-        "mn-MN-YesuiNeural",
-        "ne-NP-HemkalaNeural",
-        "ne-NP-SagarNeural",
-        "nb-NO-FinnNeural",
-        "nb-NO-PernilleNeural",
-        "ps-AF-GulNawazNeural",
-        "ps-AF-LatifaNeural",
-        "fa-IR-DilaraNeural",
-        "fa-IR-FaridNeural",
-        "pl-PL-MarekNeural",
-        "pl-PL-ZofiaNeural",
-        "pt-BR-ThalitaNeural",
-        "pt-BR-AntonioNeural",
-        "pt-BR-FranciscaNeural",
-        "pt-PT-DuarteNeural",
-        "pt-PT-RaquelNeural",
-        "ro-RO-AlinaNeural",
-        "ro-RO-EmilNeural",
-        "ru-RU-DmitryNeural",
-        "ru-RU-SvetlanaNeural",
-        "sr-RS-NicholasNeural",
-        "sr-RS-SophieNeural",
-        "si-LK-SameeraNeural",
-        "si-LK-ThiliniNeural",
-        "sk-SK-LukasNeural",
-        "sk-SK-ViktoriaNeural",
-        "sl-SI-PetraNeural",
-        "sl-SI-RokNeural",
-        "so-SO-MuuseNeural",
-        "so-SO-UbaxNeural",
-        "es-AR-ElenaNeural",
-        "es-AR-TomasNeural",
-        "es-BO-MarceloNeural",
-        "es-BO-SofiaNeural",
-        "es-CL-CatalinaNeural",
-        "es-CL-LorenzoNeural",
-        "es-ES-XimenaNeural",
-        "es-CO-GonzaloNeural",
-        "es-CO-SalomeNeural",
-        "es-CR-JuanNeural",
-        "es-CR-MariaNeural",
-        "es-CU-BelkysNeural",
-        "es-CU-ManuelNeural",
-        "es-DO-EmilioNeural",
-        "es-DO-RamonaNeural",
-        "es-EC-AndreaNeural",
-        "es-EC-LuisNeural",
-        "es-SV-LorenaNeural",
-        "es-SV-RodrigoNeural",
-        "es-GQ-JavierNeural",
-        "es-GQ-TeresaNeural",
-        "es-GT-AndresNeural",
-        "es-GT-MartaNeural",
-        "es-HN-CarlosNeural",
-        "es-HN-KarlaNeural",
-        "es-MX-DaliaNeural",
-        "es-MX-JorgeNeural",
-        "es-NI-FedericoNeural",
-        "es-NI-YolandaNeural",
-        "es-PA-MargaritaNeural",
-        "es-PA-RobertoNeural",
-        "es-PY-MarioNeural",
-        "es-PY-TaniaNeural",
-        "es-PE-AlexNeural",
-        "es-PE-CamilaNeural",
-        "es-PR-KarinaNeural",
-        "es-PR-VictorNeural",
-        "es-ES-AlvaroNeural",
-        "es-ES-ElviraNeural",
-        "es-US-AlonsoNeural",
-        "es-US-PalomaNeural",
-        "es-UY-MateoNeural",
-        "es-UY-ValentinaNeural",
-        "es-VE-PaolaNeural",
-        "es-VE-SebastianNeural",
-        "su-ID-JajangNeural",
-        "su-ID-TutiNeural",
-        "sw-KE-RafikiNeural",
-        "sw-KE-ZuriNeural",
-        "sw-TZ-DaudiNeural",
-        "sw-TZ-RehemaNeural",
-        "sv-SE-MattiasNeural",
-        "sv-SE-SofieNeural",
-        "ta-IN-PallaviNeural",
-        "ta-IN-ValluvarNeural",
-        "ta-MY-KaniNeural",
-        "ta-MY-SuryaNeural",
-        "ta-SG-AnbuNeural",
-        "ta-SG-VenbaNeural",
-        "ta-LK-KumarNeural",
-        "ta-LK-SaranyaNeural",
-        "te-IN-MohanNeural",
-        "te-IN-ShrutiNeural",
-        "th-TH-NiwatNeural",
-        "th-TH-PremwadeeNeural",
-        "tr-TR-AhmetNeural",
-        "tr-TR-EmelNeural",
-        "uk-UA-OstapNeural",
-        "uk-UA-PolinaNeural",
-        "ur-IN-GulNeural",
-        "ur-IN-SalmanNeural",
-        "ur-PK-AsadNeural",
-        "ur-PK-UzmaNeural",
-        "uz-UZ-MadinaNeural",
-        "uz-UZ-SardorNeural",
-        "vi-VN-HoaiMyNeural",
-        "vi-VN-NamMinhNeural",
-        "cy-GB-AledNeural",
-        "cy-GB-NiaNeural",
-        "zu-ZA-ThandoNeural",
-        "zu-ZA-ThembaNeural"
-    ],
-    "google_tts_voice": [
-        "af",
-        "am",
-        "ar",
-        "bg",
-        "bn",
-        "bs",
-        "ca",
-        "cs",
-        "cy",
-        "da",
-        "de",
-        "el",
-        "en",
-        "es",
-        "et",
-        "eu",
-        "fi",
-        "fr",
-        "fr-CA",
-        "gl",
-        "gu",
-        "ha",
-        "hi",
-        "hr",
-        "hu",
-        "id",
-        "is",
-        "it",
-        "iw",
-        "ja",
-        "jw",
-        "km",
-        "kn",
-        "ko",
-        "la",
-        "lt",
-        "lv",
-        "ml",
-        "mr",
-        "ms",
-        "my",
-        "ne",
-        "nl",
-        "no",
-        "pa",
-        "pl",
-        "pt",
-        "pt-PT",
-        "ro",
-        "ru",
-        "si",
-        "sk",
-        "sq",
-        "sr",
-        "su",
-        "sv",
-        "sw",
-        "ta",
-        "te",
-        "th",
-        "tl",
-        "tr",
-        "uk",
-        "ur",
-        "vi",
-        "yue",
-        "zh-CN",
-        "zh-TW",
-        "zh"
-    ],
-    "fp16": false,
-    "separator_tab": true,
-    "convert_tab": true,
-    "convert_with_whisper": true,
-    "tts_tab": true,
-    "audioldm2": true,
-    "effects_tab": true,
-    "create_dataset_tab": true,
-    "training_tab": true,
-    "fushion_tab": true,
-    "read_tab": true,
-    "onnx_tab": true,
-    "downloads_tab": true,
-    "f0_extractor_tab": true,
-    "settings_tab": true,
-    "report_bug_tab": true,
-    "font": "",
-    "app_port": 7860,
-    "tensorboard_port": 6870,
-    "num_of_restart": 5,
-    "server_name": "0.0.0.0",
-    "app_show_error": true
-}

main/configs/config.py

@@ -1,90 +0,0 @@
-import os
-import json
-import torch
-
-
-version_config_paths = [os.path.join(version, size) for version in ["v1", "v2"] for size in ["32000.json", "40000.json", "48000.json"]]
-
-def singleton(cls):
-    instances = {}
-
-    def get_instance(*args, **kwargs):
-        if cls not in instances: instances[cls] = cls(*args, **kwargs)
-        return instances[cls]
-
-    return get_instance
-
-@singleton
-class Config:
-    def __init__(self):
-        self.device = "cuda:0" if torch.cuda.is_available() else "cpu"
-        self.configs = json.load(open(os.path.join("main", "configs", "config.json"), "r"))
-        self.translations = self.multi_language()
-        self.json_config = self.load_config_json()
-        self.gpu_mem = None
-        self.per_preprocess = 3.7
-        self.is_half = self.is_fp16()
-        self.x_pad, self.x_query, self.x_center, self.x_max = self.device_config()
-    
-    def multi_language(self):
-        try:
-            lang = self.configs.get("language", "en-US")
-            if len([l for l in os.listdir(os.path.join("assets", "languages")) if l.endswith(".json")]) < 1: raise FileNotFoundError("Không tìm thấy bất cứ gói ngôn ngữ nào(No package languages found)")
-
-            if not lang: lang = "en-US"
-            if lang not in self.configs["support_language"]: raise ValueError("Language not supported....")
-
-            lang_path = os.path.join("assets", "languages", f"{lang}.json")
-            if not os.path.exists(lang_path): lang_path = os.path.join("assets", "languages", "en-US.json")
-
-            with open(lang_path, encoding="utf-8") as f:
-                translations = json.load(f)
-        except json.JSONDecodeError:
-            print(self.translations["empty_json"].format(file=lang))
-            pass
-
-        return translations
-    
-    def is_fp16(self):
-        fp16 = self.configs.get("fp16", False)
-
-        if self.device in ["cpu", "mps"] and fp16:
-            self.configs["fp16"] = False
-            fp16 = False
-
-            with open(os.path.join("main", "configs", "config.json"), "w") as f:
-                json.dump(self.configs, f, indent=4)
-        
-        if not fp16: self.preprocess_per = 3.0
-        return fp16
-
-    def load_config_json(self):
-        configs = {}
-
-        for config_file in version_config_paths:
-            try:
-                with open(os.path.join("main", "configs", config_file), "r") as f:
-                    configs[config_file] = json.load(f)
-            except json.JSONDecodeError:
-                print(self.translations["empty_json"].format(file=config_file))
-                pass
-
-        return configs
-
-    def device_config(self):
-        if self.device.startswith("cuda"): self.set_cuda_config()
-        elif self.has_mps(): self.device = "mps"
-        else: self.device = "cpu"
-
-        if self.gpu_mem is not None and self.gpu_mem <= 4: 
-            self.preprocess_per = 3.0
-            return 1, 5, 30, 32
-        
-        return (3, 10, 60, 65) if self.is_half else (1, 6, 38, 41)
-
-    def set_cuda_config(self):
-        i_device = int(self.device.split(":")[-1])
-        self.gpu_mem = torch.cuda.get_device_properties(i_device).total_memory // (1024**3)
-
-    def has_mps(self):
-        return torch.backends.mps.is_available()

main/configs/decrypt.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:330268cbf6b9317a76510b533e1640ef48ed074a07c013e5b1abc4d48cfd9dce
-size 32

main/configs/v1/32000.json

@@ -1,46 +0,0 @@
-{
-    "train": {
-        "log_interval": 200,
-        "seed": 1234,
-        "epochs": 20000,
-        "learning_rate": 0.0001,
-        "betas": [0.8, 0.99],
-        "eps": 1e-09,
-        "batch_size": 4,
-        "lr_decay": 0.999875,
-        "segment_size": 12800,
-        "init_lr_ratio": 1,
-        "warmup_epochs": 0,
-        "c_mel": 45,
-        "c_kl": 1.0
-    },
-    "data": {
-        "max_wav_value": 32768.0,
-        "sample_rate": 32000,
-        "filter_length": 1024,
-        "hop_length": 320,
-        "win_length": 1024,
-        "n_mel_channels": 80,
-        "mel_fmin": 0.0,
-        "mel_fmax": null
-    },
-    "model": {
-        "inter_channels": 192,
-        "hidden_channels": 192,
-        "filter_channels": 768,
-        "text_enc_hidden_dim": 256,
-        "n_heads": 2,
-        "n_layers": 6,
-        "kernel_size": 3,
-        "p_dropout": 0,
-        "resblock": "1",
-        "resblock_kernel_sizes": [3, 7, 11],
-        "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
-        "upsample_rates": [10, 4, 2, 2, 2],
-        "upsample_initial_channel": 512,
-        "upsample_kernel_sizes": [16, 16, 4, 4, 4],
-        "use_spectral_norm": false,
-        "gin_channels": 256,
-        "spk_embed_dim": 109
-    }
-}

main/configs/v1/40000.json

@@ -1,46 +0,0 @@
-{
-    "train": {
-        "log_interval": 200,
-        "seed": 1234,
-        "epochs": 20000,
-        "learning_rate": 0.0001,
-        "betas": [0.8, 0.99],
-        "eps": 1e-09,
-        "batch_size": 4,
-        "lr_decay": 0.999875,
-        "segment_size": 12800,
-        "init_lr_ratio": 1,
-        "warmup_epochs": 0,
-        "c_mel": 45,
-        "c_kl": 1.0
-    },
-    "data": {
-        "max_wav_value": 32768.0,
-        "sample_rate": 40000,
-        "filter_length": 2048,
-        "hop_length": 400,
-        "win_length": 2048,
-        "n_mel_channels": 125,
-        "mel_fmin": 0.0,
-        "mel_fmax": null
-    },
-    "model": {
-        "inter_channels": 192,
-        "hidden_channels": 192,
-        "filter_channels": 768,
-        "text_enc_hidden_dim": 256,
-        "n_heads": 2,
-        "n_layers": 6,
-        "kernel_size": 3,
-        "p_dropout": 0,
-        "resblock": "1",
-        "resblock_kernel_sizes": [3, 7, 11],
-        "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
-        "upsample_rates": [10, 10, 2, 2],
-        "upsample_initial_channel": 512,
-        "upsample_kernel_sizes": [16, 16, 4, 4],
-        "use_spectral_norm": false,
-        "gin_channels": 256,
-        "spk_embed_dim": 109
-    }
-}

main/configs/v1/48000.json

@@ -1,46 +0,0 @@
-{
-    "train": {
-        "log_interval": 200,
-        "seed": 1234,
-        "epochs": 20000,
-        "learning_rate": 0.0001,
-        "betas": [0.8, 0.99],
-        "eps": 1e-09,
-        "batch_size": 4,
-        "lr_decay": 0.999875,
-        "segment_size": 11520,
-        "init_lr_ratio": 1,
-        "warmup_epochs": 0,
-        "c_mel": 45,
-        "c_kl": 1.0
-    },
-    "data": {
-        "max_wav_value": 32768.0,
-        "sample_rate": 48000,
-        "filter_length": 2048,
-        "hop_length": 480,
-        "win_length": 2048,
-        "n_mel_channels": 128,
-        "mel_fmin": 0.0,
-        "mel_fmax": null
-    },
-    "model": {
-        "inter_channels": 192,
-        "hidden_channels": 192,
-        "filter_channels": 768,
-        "text_enc_hidden_dim": 256,
-        "n_heads": 2,
-        "n_layers": 6,
-        "kernel_size": 3,
-        "p_dropout": 0,
-        "resblock": "1",
-        "resblock_kernel_sizes": [3, 7, 11],
-        "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
-        "upsample_rates": [10, 6, 2, 2, 2],
-        "upsample_initial_channel": 512,
-        "upsample_kernel_sizes": [16, 16, 4, 4, 4],
-        "use_spectral_norm": false,
-        "gin_channels": 256,
-        "spk_embed_dim": 109
-    }
-}

main/configs/v2/32000.json

@@ -1,42 +0,0 @@
-{
-    "train": {
-        "log_interval": 200,
-        "seed": 1234,
-        "learning_rate": 0.0001,
-        "betas": [0.8, 0.99],
-        "eps": 1e-09,
-        "lr_decay": 0.999875,
-        "segment_size": 12800,
-        "c_mel": 45,
-        "c_kl": 1.0
-    },
-    "data": {
-        "max_wav_value": 32768.0,
-        "sample_rate": 32000,
-        "filter_length": 1024,
-        "hop_length": 320,
-        "win_length": 1024,
-        "n_mel_channels": 80,
-        "mel_fmin": 0.0,
-        "mel_fmax": null
-    },
-    "model": {
-        "inter_channels": 192,
-        "hidden_channels": 192,
-        "filter_channels": 768,
-        "text_enc_hidden_dim": 768,
-        "n_heads": 2,
-        "n_layers": 6,
-        "kernel_size": 3,
-        "p_dropout": 0,
-        "resblock": "1",
-        "resblock_kernel_sizes": [3, 7, 11],
-        "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
-        "upsample_rates": [10, 8, 2, 2],
-        "upsample_initial_channel": 512,
-        "upsample_kernel_sizes": [20, 16, 4, 4],
-        "use_spectral_norm": false,
-        "gin_channels": 256,
-        "spk_embed_dim": 109
-    }
-}

main/configs/v2/40000.json

@@ -1,42 +0,0 @@
-{
-    "train": {
-        "log_interval": 200,
-        "seed": 1234,
-        "learning_rate": 0.0001,
-        "betas": [0.8, 0.99],
-        "eps": 1e-09,
-        "lr_decay": 0.999875,
-        "segment_size": 12800,
-        "c_mel": 45,
-        "c_kl": 1.0
-    },
-    "data": {
-        "max_wav_value": 32768.0,
-        "sample_rate": 40000,
-        "filter_length": 2048,
-        "hop_length": 400,
-        "win_length": 2048,
-        "n_mel_channels": 125,
-        "mel_fmin": 0.0,
-        "mel_fmax": null
-    },
-    "model": {
-        "inter_channels": 192,
-        "hidden_channels": 192,
-        "filter_channels": 768,
-        "text_enc_hidden_dim": 768,
-        "n_heads": 2,
-        "n_layers": 6,
-        "kernel_size": 3,
-        "p_dropout": 0,
-        "resblock": "1",
-        "resblock_kernel_sizes": [3, 7, 11],
-        "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
-        "upsample_rates": [10, 10, 2, 2],
-        "upsample_initial_channel": 512,
-        "upsample_kernel_sizes": [16, 16, 4, 4],
-        "use_spectral_norm": false,
-        "gin_channels": 256,
-        "spk_embed_dim": 109
-    }
-}

main/configs/v2/48000.json

@@ -1,42 +0,0 @@
-{
-    "train": {
-        "log_interval": 200,
-        "seed": 1234,
-        "learning_rate": 0.0001,
-        "betas": [0.8, 0.99],
-        "eps": 1e-09,
-        "lr_decay": 0.999875,
-        "segment_size": 17280,
-        "c_mel": 45,
-        "c_kl": 1.0
-    },
-    "data": {
-        "max_wav_value": 32768.0,
-        "sample_rate": 48000,
-        "filter_length": 2048,
-        "hop_length": 480,
-        "win_length": 2048,
-        "n_mel_channels": 128,
-        "mel_fmin": 0.0,
-        "mel_fmax": null
-    },
-    "model": {
-        "inter_channels": 192,
-        "hidden_channels": 192,
-        "filter_channels": 768,
-        "text_enc_hidden_dim": 768,
-        "n_heads": 2,
-        "n_layers": 6,
-        "kernel_size": 3,
-        "p_dropout": 0,
-        "resblock": "1",
-        "resblock_kernel_sizes": [3, 7, 11],
-        "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
-        "upsample_rates": [12, 10, 2, 2],
-        "upsample_initial_channel": 512,
-        "upsample_kernel_sizes": [24, 20, 4, 4],
-        "use_spectral_norm": false,
-        "gin_channels": 256,
-        "spk_embed_dim": 109
-    }
-}

main/inference/audio_effects.py

@@ -1,180 +0,0 @@
-import os
-import sys
-import librosa
-import argparse
-
-import numpy as np
-import soundfile as sf
-
-from distutils.util import strtobool
-from scipy.signal import butter, filtfilt
-from pedalboard import Pedalboard, Chorus, Distortion, Reverb, PitchShift, Delay, Limiter, Gain, Bitcrush, Clipping, Compressor, Phaser, HighpassFilter
-
-sys.path.append(os.getcwd())
-
-from main.configs.config import Config
-from main.library.utils import pydub_convert, pydub_load
-
-translations = Config().translations
-
-def parse_arguments():
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--input_path", type=str, required=True)
-    parser.add_argument("--output_path", type=str, default="./audios/apply_effects.wav")
-    parser.add_argument("--export_format", type=str, default="wav")
-    parser.add_argument("--resample", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--resample_sr", type=int, default=0)
-    parser.add_argument("--chorus", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--chorus_depth", type=float, default=0.5)
-    parser.add_argument("--chorus_rate", type=float, default=1.5)
-    parser.add_argument("--chorus_mix", type=float, default=0.5)
-    parser.add_argument("--chorus_delay", type=int, default=10)
-    parser.add_argument("--chorus_feedback", type=float, default=0)
-    parser.add_argument("--distortion", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--drive_db", type=int, default=20)
-    parser.add_argument("--reverb", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--reverb_room_size", type=float, default=0.5)
-    parser.add_argument("--reverb_damping", type=float, default=0.5)
-    parser.add_argument("--reverb_wet_level", type=float, default=0.33)
-    parser.add_argument("--reverb_dry_level", type=float, default=0.67)
-    parser.add_argument("--reverb_width", type=float, default=1)
-    parser.add_argument("--reverb_freeze_mode", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--pitchshift", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--pitch_shift", type=int, default=0)
-    parser.add_argument("--delay", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--delay_seconds", type=float, default=0.5)
-    parser.add_argument("--delay_feedback", type=float, default=0.5)
-    parser.add_argument("--delay_mix", type=float, default=0.5)
-    parser.add_argument("--compressor", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--compressor_threshold", type=int, default=-20)
-    parser.add_argument("--compressor_ratio", type=float, default=4)
-    parser.add_argument("--compressor_attack_ms", type=float, default=10)
-    parser.add_argument("--compressor_release_ms", type=int, default=200)
-    parser.add_argument("--limiter", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--limiter_threshold", type=int, default=0)
-    parser.add_argument("--limiter_release", type=int, default=100)
-    parser.add_argument("--gain", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--gain_db", type=int, default=0)
-    parser.add_argument("--bitcrush", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--bitcrush_bit_depth", type=int, default=16)
-    parser.add_argument("--clipping", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--clipping_threshold", type=int, default=-10)
-    parser.add_argument("--phaser", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--phaser_rate_hz", type=float, default=0.5)
-    parser.add_argument("--phaser_depth", type=float, default=0.5)
-    parser.add_argument("--phaser_centre_frequency_hz", type=int, default=1000)
-    parser.add_argument("--phaser_feedback", type=float, default=0)
-    parser.add_argument("--phaser_mix", type=float, default=0.5)
-    parser.add_argument("--treble_bass_boost", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--bass_boost_db", type=int, default=0)
-    parser.add_argument("--bass_boost_frequency", type=int, default=100)
-    parser.add_argument("--treble_boost_db", type=int, default=0)
-    parser.add_argument("--treble_boost_frequency", type=int, default=3000)
-    parser.add_argument("--fade_in_out", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--fade_in_duration", type=float, default=2000)
-    parser.add_argument("--fade_out_duration", type=float, default=2000)
-    parser.add_argument("--audio_combination", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--audio_combination_input", type=str)
-    
-    return parser.parse_args()
-
-def process_audio(input_path, output_path, resample, resample_sr, chorus_depth, chorus_rate, chorus_mix, chorus_delay, chorus_feedback, distortion_drive, reverb_room_size, reverb_damping, reverb_wet_level, reverb_dry_level, reverb_width, reverb_freeze_mode, pitch_shift, delay_seconds, delay_feedback, delay_mix, compressor_threshold, compressor_ratio, compressor_attack_ms, compressor_release_ms, limiter_threshold, limiter_release, gain_db, bitcrush_bit_depth, clipping_threshold, phaser_rate_hz, phaser_depth, phaser_centre_frequency_hz, phaser_feedback, phaser_mix, bass_boost_db, bass_boost_frequency, treble_boost_db, treble_boost_frequency, fade_in_duration, fade_out_duration, export_format, chorus, distortion, reverb, pitchshift, delay, compressor, limiter, gain, bitcrush, clipping, phaser, treble_bass_boost, fade_in_out, audio_combination, audio_combination_input):
-    def bass_boost(audio, gain_db, frequency, sample_rate):
-        if gain_db >= 1:
-            b, a = butter(4, frequency / (0.5 * sample_rate), btype='low')
-
-            return filtfilt(b, a, audio) * 10 ** (gain_db / 20)
-        else: return audio
-    
-    def treble_boost(audio, gain_db, frequency, sample_rate):
-        if gain_db >=1:
-            b, a = butter(4, frequency / (0.5 * sample_rate), btype='high')
-            
-            return filtfilt(b, a, audio) * 10 ** (gain_db / 20)
-        else: return audio
-
-    def fade_out_effect(audio, sr, duration=3.0):
-        length = int(duration * sr)
-        end = audio.shape[0]
-        
-        if length > end: length = end  
-        start = end - length
-
-        audio[start:end] = audio[start:end] * np.linspace(1.0, 0.0, length)
-        return audio
-
-    def fade_in_effect(audio, sr, duration=3.0):
-        length = int(duration * sr)
-        start = 0
-
-        if length > audio.shape[0]: length = audio.shape[0]  
-        end = length
-        
-        audio[start:end] = audio[start:end] * np.linspace(0.0, 1.0, length)
-        return audio
-
-    if not input_path or not os.path.exists(input_path): 
-        print(translations["input_not_valid"])
-        sys.exit(1)
-
-    if not output_path: 
-        print(translations["output_not_valid"])
-        sys.exit(1)
-    
-    if os.path.exists(output_path): os.remove(output_path)
-    
-    try:
-        input_path = input_path.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
-        
-        try:
-            audio, sample_rate = sf.read(input_path, dtype=np.float32)
-        except:
-            audio, sample_rate = librosa.load(input_path, sr=None)
-    except Exception as e:
-        raise RuntimeError(f"{translations['errors_loading_audio']}: {e}")
-    
-    audio = audio.flatten()
-
-    try:
-        board = Pedalboard([HighpassFilter()])
-
-        if chorus: board.append(Chorus(depth=chorus_depth, rate_hz=chorus_rate, mix=chorus_mix, centre_delay_ms=chorus_delay, feedback=chorus_feedback))
-        if distortion: board.append(Distortion(drive_db=distortion_drive))
-        if reverb: board.append(Reverb(room_size=reverb_room_size, damping=reverb_damping, wet_level=reverb_wet_level, dry_level=reverb_dry_level, width=reverb_width, freeze_mode=1 if reverb_freeze_mode else 0))
-        if pitchshift: board.append(PitchShift(semitones=pitch_shift))
-        if delay: board.append(Delay(delay_seconds=delay_seconds, feedback=delay_feedback, mix=delay_mix))
-        if compressor: board.append(Compressor(threshold_db=compressor_threshold, ratio=compressor_ratio, attack_ms=compressor_attack_ms, release_ms=compressor_release_ms))
-        if limiter: board.append(Limiter(threshold_db=limiter_threshold, release_ms=limiter_release))
-        if gain: board.append(Gain(gain_db=gain_db))
-        if bitcrush: board.append(Bitcrush(bit_depth=bitcrush_bit_depth))
-        if clipping: board.append(Clipping(threshold_db=clipping_threshold)) 
-        if phaser: board.append(Phaser(rate_hz=phaser_rate_hz, depth=phaser_depth, centre_frequency_hz=phaser_centre_frequency_hz, feedback=phaser_feedback, mix=phaser_mix))
-
-        processed_audio = board(audio, sample_rate)
-
-        if treble_bass_boost:
-            processed_audio = bass_boost(processed_audio, bass_boost_db, bass_boost_frequency, sample_rate)
-            processed_audio = treble_boost(processed_audio, treble_boost_db, treble_boost_frequency, sample_rate)
-
-        if fade_in_out:
-            processed_audio = fade_in_effect(processed_audio, sample_rate, fade_in_duration)
-            processed_audio = fade_out_effect(processed_audio, sample_rate, fade_out_duration)
-            
-        if resample_sr != sample_rate and resample_sr > 0 and resample:
-            target_sr = min([8000, 11025, 12000, 16000, 22050, 24000, 32000, 44100, 48000, 96000], key=lambda x: abs(x - resample_sr))
-            processed_audio = librosa.resample(processed_audio, orig_sr=sample_rate, target_sr=target_sr, res_type="soxr_vhq")
-            sample_rate = target_sr
-
-        sf.write(output_path.replace("wav", export_format), processed_audio, sample_rate, format=export_format)
-
-        if audio_combination: pydub_convert(pydub_load(audio_combination_input)).overlay(pydub_convert(pydub_load(output_path.replace("wav", export_format)))).export(output_path.replace("wav", export_format), format=export_format)
-    except Exception as e:
-        raise RuntimeError(translations["apply_error"].format(e=e))
-    
-    return output_path
-
-def main():
-    args = parse_arguments()
-    process_audio(input_path=args.input_path, output_path=args.output_path, resample=args.resample, resample_sr=args.resample_sr, chorus_depth=args.chorus_depth, chorus_rate=args.chorus_rate, chorus_mix=args.chorus_mix, chorus_delay=args.chorus_delay, chorus_feedback=args.chorus_feedback, distortion_drive=args.drive_db, reverb_room_size=args.reverb_room_size, reverb_damping=args.reverb_damping, reverb_wet_level=args.reverb_wet_level, reverb_dry_level=args.reverb_dry_level, reverb_width=args.reverb_width, reverb_freeze_mode=args.reverb_freeze_mode, pitch_shift=args.pitch_shift, delay_seconds=args.delay_seconds, delay_feedback=args.delay_feedback, delay_mix=args.delay_mix, compressor_threshold=args.compressor_threshold, compressor_ratio=args.compressor_ratio, compressor_attack_ms=args.compressor_attack_ms, compressor_release_ms=args.compressor_release_ms, limiter_threshold=args.limiter_threshold, limiter_release=args.limiter_release, gain_db=args.gain_db, bitcrush_bit_depth=args.bitcrush_bit_depth, clipping_threshold=args.clipping_threshold, phaser_rate_hz=args.phaser_rate_hz, phaser_depth=args.phaser_depth, phaser_centre_frequency_hz=args.phaser_centre_frequency_hz, phaser_feedback=args.phaser_feedback, phaser_mix=args.phaser_mix, bass_boost_db=args.bass_boost_db, bass_boost_frequency=args.bass_boost_frequency, treble_boost_db=args.treble_boost_db, treble_boost_frequency=args.treble_boost_frequency, fade_in_duration=args.fade_in_duration, fade_out_duration=args.fade_out_duration, export_format=args.export_format, chorus=args.chorus, distortion=args.distortion, reverb=args.reverb, pitchshift=args.pitchshift, delay=args.delay, compressor=args.compressor, limiter=args.limiter, gain=args.gain, bitcrush=args.bitcrush, clipping=args.clipping, phaser=args.phaser, treble_bass_boost=args.treble_bass_boost, fade_in_out=args.fade_in_out, audio_combination=args.audio_combination, audio_combination_input=args.audio_combination_input)
-
-if __name__ == "__main__": main()

main/inference/audioldm2.py

@@ -1,210 +0,0 @@
-import os
-import sys
-import time
-import tqdm
-import torch
-import logging
-import librosa
-import argparse
-import scipy.signal
-import logging.handlers
-
-import numpy as np
-import soundfile as sf
-
-from torch import inference_mode
-from distutils.util import strtobool
-
-sys.path.append(os.getcwd())
-
-from main.configs.config import Config
-from main.library.audioldm2.utils import load_audio
-from main.library.audioldm2.models import load_model
-
-config = Config()
-translations = config.translations
-logger = logging.getLogger(__name__)
-logger.propagate = False
-
-for l in ["torch", "httpx", "httpcore", "diffusers", "transformers"]:
-    logging.getLogger(l).setLevel(logging.ERROR)
-
-if logger.hasHandlers(): logger.handlers.clear()
-else:
-    console_handler = logging.StreamHandler()
-    console_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
-    console_handler.setFormatter(console_formatter)
-    console_handler.setLevel(logging.INFO)
-    file_handler = logging.handlers.RotatingFileHandler(os.path.join("assets", "logs", "audioldm2.log"), maxBytes=5*1024*1024, backupCount=3, encoding='utf-8')
-    file_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
-    file_handler.setFormatter(file_formatter)
-    file_handler.setLevel(logging.DEBUG)
-    logger.addHandler(console_handler)
-    logger.addHandler(file_handler)
-    logger.setLevel(logging.DEBUG)
-
-def parse_arguments():
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--input_path", type=str, required=True)
-    parser.add_argument("--output_path", type=str, default="./output.wav")
-    parser.add_argument("--export_format", type=str, default="wav")
-    parser.add_argument("--sample_rate", type=int, default=44100)
-    parser.add_argument("--audioldm_model", type=str, default="audioldm2-music")
-    parser.add_argument("--source_prompt", type=str, default="")
-    parser.add_argument("--target_prompt", type=str, default="")
-    parser.add_argument("--steps", type=int, default=200)
-    parser.add_argument("--cfg_scale_src", type=float, default=3.5)
-    parser.add_argument("--cfg_scale_tar", type=float, default=12)
-    parser.add_argument("--t_start", type=int, default=45)
-    parser.add_argument("--save_compute", type=lambda x: bool(strtobool(x)), default=False)
-
-    return parser.parse_args()
-
-def main():
-    args = parse_arguments()
-    input_path, output_path, export_format, sample_rate, audioldm_model, source_prompt, target_prompt, steps, cfg_scale_src, cfg_scale_tar, t_start, save_compute = args.input_path, args.output_path, args.export_format, args.sample_rate, args.audioldm_model, args.source_prompt, args.target_prompt, args.steps, args.cfg_scale_src, args.cfg_scale_tar, args.t_start, args.save_compute
-
-    log_data = {translations['audio_path']: input_path, translations['output_path']: output_path.replace('wav', export_format), translations['model_name']: audioldm_model, translations['export_format']: export_format, translations['sample_rate']: sample_rate, translations['steps']: steps, translations['source_prompt']: source_prompt, translations['target_prompt']: target_prompt, translations['cfg_scale_src']: cfg_scale_src, translations['cfg_scale_tar']: cfg_scale_tar, translations['t_start']: t_start, translations['save_compute']: save_compute}
-
-    for key, value in log_data.items():
-        logger.debug(f"{key}: {value}")
-   
-    start_time = time.time()
-    logger.info(translations["start_edit"].format(input_path=input_path))
-    pid_path = os.path.join("assets", "audioldm2_pid.txt")
-    with open(pid_path, "w") as pid_file:
-        pid_file.write(str(os.getpid()))
-    
-    try:
-        edit(input_path, output_path, audioldm_model, source_prompt, target_prompt, steps, cfg_scale_src, cfg_scale_tar, t_start, save_compute, sample_rate, config.device, export_format=export_format)
-    except Exception as e:
-        logger.error(translations["error_edit"].format(e=e))
-        import traceback
-        logger.debug(traceback.format_exc())
-        
-    logger.info(translations["edit_success"].format(time=f"{(time.time() - start_time):.2f}", output_path=output_path.replace('wav', export_format)))
-
-def invert(ldm_stable, x0, prompt_src, num_diffusion_steps, cfg_scale_src, duration, save_compute):
-    with inference_mode():
-        w0 = ldm_stable.vae_encode(x0)
-
-    _, zs, wts, extra_info = inversion_forward_process(ldm_stable, w0, etas=1, prompts=[prompt_src], cfg_scales=[cfg_scale_src], num_inference_steps=num_diffusion_steps, numerical_fix=True, duration=duration, save_compute=save_compute)
-    return zs, wts, extra_info
-
-def low_pass_filter(audio, cutoff=7500, sr=16000):
-    b, a = scipy.signal.butter(4, cutoff / (sr / 2), btype='low')
-    return scipy.signal.filtfilt(b, a, audio)
-
-def sample(output_audio, sr, ldm_stable, zs, wts, extra_info, prompt_tar, tstart, cfg_scale_tar, duration, save_compute, export_format = "wav"):
-    tstart = torch.tensor(tstart, dtype=torch.int32)
-    w0, _ = inversion_reverse_process(ldm_stable, xT=wts, tstart=tstart, etas=1., prompts=[prompt_tar], neg_prompts=[""], cfg_scales=[cfg_scale_tar], zs=zs[:int(tstart)], duration=duration, extra_info=extra_info, save_compute=save_compute)
-
-    with inference_mode():
-        x0_dec = ldm_stable.vae_decode(w0.to(torch.float16 if config.is_half else torch.float32))
-
-    if x0_dec.dim() < 4: x0_dec = x0_dec[None, :, :, :]
-
-    with torch.no_grad():
-        audio = ldm_stable.decode_to_mel(x0_dec.to(torch.float16 if config.is_half else torch.float32))
-
-    audio = audio.float().squeeze().cpu().numpy()
-    orig_sr = 16000
-
-    if sr != 16000 and sr > 0: 
-        audio = librosa.resample(audio, orig_sr=orig_sr, target_sr=sr, res_type="soxr_vhq")
-        orig_sr = sr
-
-    audio = low_pass_filter(audio, 7500, orig_sr)
-
-    sf.write(output_audio, np.tile(audio, (2, 1)).T, orig_sr, format=export_format)
-    return output_audio
-
-def edit(input_audio, output_audio, model_id, source_prompt = "", target_prompt = "", steps = 200, cfg_scale_src = 3.5, cfg_scale_tar = 12, t_start = 45, save_compute = True, sr = 44100, device = "cpu", export_format = "wav"):
-    ldm_stable = load_model(model_id, device=device)
-    ldm_stable.model.scheduler.set_timesteps(steps, device=device)
-    x0, duration = load_audio(input_audio, ldm_stable.get_melspectrogram(), device=device)
-    zs_tensor, wts_tensor, extra_info_list = invert(ldm_stable=ldm_stable, x0=x0, prompt_src=source_prompt, num_diffusion_steps=steps, cfg_scale_src=cfg_scale_src, duration=duration, save_compute=save_compute)
-
-    return sample(output_audio, sr, ldm_stable, zs_tensor, wts_tensor, extra_info_list, prompt_tar=target_prompt, tstart=int(t_start / 100 * steps), cfg_scale_tar=cfg_scale_tar, duration=duration, save_compute=save_compute, export_format=export_format)
-
-def inversion_forward_process(model, x0, etas = None, prompts = [""], cfg_scales = [3.5], num_inference_steps = 50, numerical_fix = False, duration = None, first_order = False, save_compute = True):
-    if len(prompts) > 1 or prompts[0] != "":
-        text_embeddings_hidden_states, text_embeddings_class_labels, text_embeddings_boolean_prompt_mask = model.encode_text(prompts)
-        uncond_embeddings_hidden_states, uncond_embeddings_class_lables, uncond_boolean_prompt_mask = model.encode_text([""], negative=True, save_compute=save_compute, cond_length=text_embeddings_class_labels.shape[1] if text_embeddings_class_labels is not None else None)
-    else: uncond_embeddings_hidden_states, uncond_embeddings_class_lables, uncond_boolean_prompt_mask = model.encode_text([""], negative=True, save_compute=False)
-
-    timesteps = model.model.scheduler.timesteps.to(model.device)
-    variance_noise_shape = model.get_noise_shape(x0, num_inference_steps)
-
-    if type(etas) in [int, float]: etas = [etas]*model.model.scheduler.num_inference_steps
-
-    xts = model.sample_xts_from_x0(x0, num_inference_steps=num_inference_steps)
-    zs = torch.zeros(size=variance_noise_shape, device=model.device)
-    extra_info = [None] * len(zs)
-
-    if timesteps[0].dtype == torch.int64: t_to_idx = {int(v): k for k, v in enumerate(timesteps)}
-    elif timesteps[0].dtype == torch.float32: t_to_idx = {float(v): k for k, v in enumerate(timesteps)}
-
-    xt = x0
-    model.setup_extra_inputs(xt, init_timestep=timesteps[0], audio_end_in_s=duration, save_compute=save_compute and prompts[0] != "")
-
-    for t in tqdm.tqdm(timesteps, desc=translations["inverting"], ncols=100, unit="a"):
-        idx = num_inference_steps - t_to_idx[int(t) if timesteps[0].dtype == torch.int64 else float(t)] - 1
-        xt = xts[idx + 1][None]
-        xt_inp = model.model.scheduler.scale_model_input(xt, t).to(torch.float16 if config.is_half else torch.float32)
-
-        with torch.no_grad():
-            if save_compute and prompts[0] != "":
-                comb_out, _, _ = model.unet_forward(xt_inp.expand(2, -1, -1, -1) if hasattr(model.model, 'unet') else xt_inp.expand(2, -1, -1), timestep=t, encoder_hidden_states=torch.cat([uncond_embeddings_hidden_states, text_embeddings_hidden_states], dim=0) if uncond_embeddings_hidden_states is not None else None, class_labels=torch.cat([uncond_embeddings_class_lables, text_embeddings_class_labels], dim=0) if uncond_embeddings_class_lables is not None else None, encoder_attention_mask=torch.cat([uncond_boolean_prompt_mask, text_embeddings_boolean_prompt_mask], dim=0) if uncond_boolean_prompt_mask is not None else None)
-                out, cond_out = comb_out.sample.chunk(2, dim=0)
-            else:
-                out = model.unet_forward(xt_inp, timestep=t, encoder_hidden_states=uncond_embeddings_hidden_states, class_labels=uncond_embeddings_class_lables, encoder_attention_mask=uncond_boolean_prompt_mask)[0].sample
-                if len(prompts) > 1 or prompts[0] != "": cond_out = model.unet_forward(xt_inp, timestep=t, encoder_hidden_states=text_embeddings_hidden_states, class_labels=text_embeddings_class_labels, encoder_attention_mask=text_embeddings_boolean_prompt_mask)[0].sample
-
-        if len(prompts) > 1 or prompts[0] != "": noise_pred = out + (cfg_scales[0] * (cond_out - out)).sum(axis=0).unsqueeze(0)
-        else: noise_pred = out
-
-        xtm1 = xts[idx][None]
-        z, xtm1, extra = model.get_zs_from_xts(xt, xtm1, noise_pred, t, eta=etas[idx], numerical_fix=numerical_fix, first_order=first_order)
-        zs[idx] = z
-        xts[idx] = xtm1
-        extra_info[idx] = extra
-
-    if zs is not None: zs[0] = torch.zeros_like(zs[0])
-    return xt, zs, xts, extra_info
-
-def inversion_reverse_process(model, xT, tstart, etas = 0, prompts = [""], neg_prompts = [""], cfg_scales = None, zs = None, duration = None, first_order = False, extra_info = None, save_compute = True):
-    text_embeddings_hidden_states, text_embeddings_class_labels, text_embeddings_boolean_prompt_mask = model.encode_text(prompts)
-    uncond_embeddings_hidden_states, uncond_embeddings_class_lables, uncond_boolean_prompt_mask = model.encode_text(neg_prompts, negative=True, save_compute=save_compute, cond_length=text_embeddings_class_labels.shape[1] if text_embeddings_class_labels is not None else None)
-    xt = xT[tstart.max()].unsqueeze(0)
-
-    if etas is None: etas = 0
-    if type(etas) in [int, float]: etas = [etas]*model.model.scheduler.num_inference_steps
-    
-    assert len(etas) == model.model.scheduler.num_inference_steps
-    timesteps = model.model.scheduler.timesteps.to(model.device)
-
-    if timesteps[0].dtype == torch.int64: t_to_idx = {int(v): k for k, v in enumerate(timesteps[-zs.shape[0]:])}
-    elif timesteps[0].dtype == torch.float32: t_to_idx = {float(v): k for k, v in enumerate(timesteps[-zs.shape[0]:])}
-
-    model.setup_extra_inputs(xt, extra_info=extra_info, init_timestep=timesteps[-zs.shape[0]], audio_end_in_s=duration, save_compute=save_compute)
-
-    for t in tqdm.tqdm(timesteps[-zs.shape[0]:], desc=translations["editing"], ncols=100, unit="a"):
-        idx = model.model.scheduler.num_inference_steps - t_to_idx[int(t) if timesteps[0].dtype == torch.int64 else float(t)] - (model.model.scheduler.num_inference_steps - zs.shape[0] + 1)
-        xt_inp = model.model.scheduler.scale_model_input(xt, t).to(torch.float16 if config.is_half else torch.float32)
-
-        with torch.no_grad():
-            if save_compute:
-                comb_out, _, _ = model.unet_forward(xt_inp.expand(2, -1, -1, -1) if hasattr(model.model, 'unet') else xt_inp.expand(2, -1, -1), timestep=t, encoder_hidden_states=torch.cat([uncond_embeddings_hidden_states, text_embeddings_hidden_states], dim=0) if uncond_embeddings_hidden_states is not None else None, class_labels=torch.cat([uncond_embeddings_class_lables, text_embeddings_class_labels], dim=0) if uncond_embeddings_class_lables is not None else None, encoder_attention_mask=torch.cat([uncond_boolean_prompt_mask, text_embeddings_boolean_prompt_mask], dim=0) if uncond_boolean_prompt_mask is not None else None)
-                uncond_out, cond_out = comb_out.sample.chunk(2, dim=0)
-            else:
-                uncond_out = model.unet_forward(xt_inp, timestep=t, encoder_hidden_states=uncond_embeddings_hidden_states, class_labels=uncond_embeddings_class_lables, encoder_attention_mask=uncond_boolean_prompt_mask)[0].sample
-                cond_out = model.unet_forward(xt_inp, timestep=t, encoder_hidden_states=text_embeddings_hidden_states, class_labels=text_embeddings_class_labels, encoder_attention_mask=text_embeddings_boolean_prompt_mask)[0].sample
-
-        z = zs[idx] if zs is not None else None
-        noise_pred = uncond_out + (cfg_scales[0] * (cond_out - uncond_out)).sum(axis=0).unsqueeze(0)
-        xt = model.reverse_step_with_custom_noise(noise_pred, t, xt, variance_noise=z.unsqueeze(0), eta=etas[idx], first_order=first_order)
-
-    return xt, zs
-
-if __name__ == "__main__": main()

main/inference/convert.py

@@ -1,590 +0,0 @@
-import re
-import os
-import gc
-import sys
-import time
-import faiss
-import torch
-import librosa
-import logging
-import argparse
-import warnings
-import onnxruntime
-import logging.handlers
-
-import numpy as np
-import soundfile as sf
-import torch.nn.functional as F
-
-from tqdm import tqdm
-from scipy import signal
-from distutils.util import strtobool
-
-warnings.filterwarnings("ignore")
-sys.path.append(os.getcwd())
-
-from main.configs.config import Config
-from main.library.algorithm.synthesizers import Synthesizer
-from main.library.utils import check_predictors, check_embedders, load_audio, load_embedders_model, cut, restore
-
-bh, ah = signal.butter(N=5, Wn=48, btype="high", fs=16000)
-config = Config()
-translations = config.translations
-logger = logging.getLogger(__name__)
-logger.propagate = False
-
-for l in ["torch", "faiss", "httpx", "httpcore", "faiss.loader", "numba.core", "urllib3", "transformers", "matplotlib"]:
-    logging.getLogger(l).setLevel(logging.ERROR)
-
-if logger.hasHandlers(): logger.handlers.clear()
-else:
-    console_handler = logging.StreamHandler()
-    console_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
-    console_handler.setFormatter(console_formatter)
-    console_handler.setLevel(logging.INFO)
-    file_handler = logging.handlers.RotatingFileHandler(os.path.join("assets", "logs", "convert.log"), maxBytes=5*1024*1024, backupCount=3, encoding='utf-8')
-    file_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
-    file_handler.setFormatter(file_formatter)
-    file_handler.setLevel(logging.DEBUG)
-    logger.addHandler(console_handler)
-    logger.addHandler(file_handler)
-    logger.setLevel(logging.DEBUG)
-
-def parse_arguments():
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--pitch", type=int, default=0)
-    parser.add_argument("--filter_radius", type=int, default=3)
-    parser.add_argument("--index_rate", type=float, default=0.5)
-    parser.add_argument("--volume_envelope", type=float, default=1)
-    parser.add_argument("--protect", type=float, default=0.33)
-    parser.add_argument("--hop_length", type=int, default=64)
-    parser.add_argument("--f0_method", type=str, default="rmvpe")
-    parser.add_argument("--embedder_model", type=str, default="contentvec_base")
-    parser.add_argument("--input_path", type=str, required=True)
-    parser.add_argument("--output_path", type=str, default="./audios/output.wav")
-    parser.add_argument("--export_format", type=str, default="wav")
-    parser.add_argument("--pth_path",  type=str,  required=True)
-    parser.add_argument("--index_path", type=str)
-    parser.add_argument("--f0_autotune", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--f0_autotune_strength", type=float, default=1)
-    parser.add_argument("--clean_audio", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--clean_strength", type=float, default=0.7)
-    parser.add_argument("--resample_sr", type=int, default=0)
-    parser.add_argument("--split_audio", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--checkpointing", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--f0_file", type=str, default="")
-    parser.add_argument("--f0_onnx", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--embedders_mode", type=str, default="fairseq")
-    parser.add_argument("--formant_shifting", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--formant_qfrency", type=float, default=0.8)
-    parser.add_argument("--formant_timbre", type=float, default=0.8)
-
-    return parser.parse_args()
-
-def main():
-    args = parse_arguments()
-    pitch, filter_radius, index_rate, volume_envelope, protect, hop_length, f0_method, input_path, output_path, pth_path, index_path, f0_autotune, f0_autotune_strength, clean_audio, clean_strength, export_format, embedder_model, resample_sr, split_audio, checkpointing, f0_file, f0_onnx, embedders_mode, formant_shifting, formant_qfrency, formant_timbre = args.pitch, args.filter_radius, args.index_rate, args.volume_envelope,args.protect, args.hop_length, args.f0_method, args.input_path, args.output_path, args.pth_path, args.index_path, args.f0_autotune, args.f0_autotune_strength, args.clean_audio, args.clean_strength, args.export_format, args.embedder_model, args.resample_sr, args.split_audio, args.checkpointing, args.f0_file, args.f0_onnx, args.embedders_mode, args.formant_shifting, args.formant_qfrency, args.formant_timbre
-
-    log_data = {translations['pitch']: pitch, translations['filter_radius']: filter_radius, translations['index_strength']: index_rate, translations['volume_envelope']: volume_envelope, translations['protect']: protect, "Hop length": hop_length, translations['f0_method']: f0_method, translations['audio_path']: input_path, translations['output_path']: output_path.replace('wav', export_format), translations['model_path']: pth_path, translations['indexpath']: index_path, translations['autotune']: f0_autotune, translations['clear_audio']: clean_audio, translations['export_format']: export_format, translations['hubert_model']: embedder_model, translations['split_audio']: split_audio, translations['memory_efficient_training']: checkpointing, translations["f0_onnx_mode"]: f0_onnx, translations["embed_mode"]: embedders_mode}
-
-    if clean_audio: log_data[translations['clean_strength']] = clean_strength
-    if resample_sr != 0: log_data[translations['sample_rate']] = resample_sr
-
-    if f0_autotune: log_data[translations['autotune_rate_info']] = f0_autotune_strength
-    if os.path.isfile(f0_file): log_data[translations['f0_file']] = f0_file
-
-    if formant_shifting:
-        log_data[translations['formant_qfrency']] = formant_qfrency
-        log_data[translations['formant_timbre']] = formant_timbre
-
-    for key, value in log_data.items():
-        logger.debug(f"{key}: {value}")
-    
-    run_convert_script(pitch=pitch, filter_radius=filter_radius, index_rate=index_rate, volume_envelope=volume_envelope, protect=protect, hop_length=hop_length, f0_method=f0_method, input_path=input_path, output_path=output_path, pth_path=pth_path, index_path=index_path, f0_autotune=f0_autotune, f0_autotune_strength=f0_autotune_strength, clean_audio=clean_audio, clean_strength=clean_strength, export_format=export_format, embedder_model=embedder_model, resample_sr=resample_sr, split_audio=split_audio, checkpointing=checkpointing, f0_file=f0_file, f0_onnx=f0_onnx, embedders_mode=embedders_mode, formant_shifting=formant_shifting, formant_qfrency=formant_qfrency, formant_timbre=formant_timbre)
-
-def run_convert_script(pitch=0, filter_radius=3, index_rate=0.5, volume_envelope=1, protect=0.5, hop_length=64, f0_method="rmvpe", input_path=None, output_path="./output.wav", pth_path=None, index_path=None, f0_autotune=False, f0_autotune_strength=1, clean_audio=False, clean_strength=0.7, export_format="wav", embedder_model="contentvec_base", resample_sr=0, split_audio=False, checkpointing=False, f0_file=None, f0_onnx=False, embedders_mode="fairseq", formant_shifting=False, formant_qfrency=0.8, formant_timbre=0.8):
-    check_predictors(f0_method, f0_onnx); check_embedders(embedder_model, embedders_mode)
-
-    if not pth_path or not os.path.exists(pth_path) or os.path.isdir(pth_path) or not pth_path.endswith((".pth", ".onnx")):
-        logger.warning(translations["provide_file"].format(filename=translations["model"]))
-        sys.exit(1)
-
-    cvt = VoiceConverter(pth_path, 0)
-    start_time = time.time()
-
-    pid_path = os.path.join("assets", "convert_pid.txt")
-    with open(pid_path, "w") as pid_file:
-        pid_file.write(str(os.getpid()))
-
-    if os.path.isdir(input_path):
-        logger.info(translations["convert_batch"])
-        audio_files = [f for f in os.listdir(input_path) if f.lower().endswith(("wav", "mp3", "flac", "ogg", "opus", "m4a", "mp4", "aac", "alac", "wma", "aiff", "webm", "ac3"))]
-
-        if not audio_files: 
-            logger.warning(translations["not_found_audio"])
-            sys.exit(1)
-
-        logger.info(translations["found_audio"].format(audio_files=len(audio_files)))
-
-        for audio in audio_files:
-            audio_path = os.path.join(input_path, audio)
-            output_audio = os.path.join(input_path, os.path.splitext(audio)[0] + f"_output.{export_format}")
-
-            logger.info(f"{translations['convert_audio']} '{audio_path}'...")
-            if os.path.exists(output_audio): os.remove(output_audio)
-            cvt.convert_audio(pitch=pitch, filter_radius=filter_radius, index_rate=index_rate, volume_envelope=volume_envelope, protect=protect, hop_length=hop_length, f0_method=f0_method, audio_input_path=audio_path, audio_output_path=output_audio, index_path=index_path, f0_autotune=f0_autotune, f0_autotune_strength=f0_autotune_strength, clean_audio=clean_audio, clean_strength=clean_strength, export_format=export_format, embedder_model=embedder_model, resample_sr=resample_sr, checkpointing=checkpointing, f0_file=f0_file, f0_onnx=f0_onnx, embedders_mode=embedders_mode, formant_shifting=formant_shifting, formant_qfrency=formant_qfrency, formant_timbre=formant_timbre, split_audio=split_audio)
-
-        logger.info(translations["convert_batch_success"].format(elapsed_time=f"{(time.time() - start_time):.2f}", output_path=output_path.replace('wav', export_format)))
-    else:
-        if not os.path.exists(input_path):
-            logger.warning(translations["not_found_audio"])
-            sys.exit(1)
-
-        logger.info(f"{translations['convert_audio']} '{input_path}'...")
-        if os.path.exists(output_path): os.remove(output_path)
-        cvt.convert_audio(pitch=pitch, filter_radius=filter_radius, index_rate=index_rate, volume_envelope=volume_envelope, protect=protect, hop_length=hop_length, f0_method=f0_method, audio_input_path=input_path, audio_output_path=output_path, index_path=index_path, f0_autotune=f0_autotune, f0_autotune_strength=f0_autotune_strength, clean_audio=clean_audio, clean_strength=clean_strength, export_format=export_format, embedder_model=embedder_model, resample_sr=resample_sr, checkpointing=checkpointing, f0_file=f0_file, f0_onnx=f0_onnx, embedders_mode=embedders_mode, formant_shifting=formant_shifting, formant_qfrency=formant_qfrency, formant_timbre=formant_timbre, split_audio=split_audio)
-
-        if os.path.exists(pid_path): os.remove(pid_path)
-        logger.info(translations["convert_audio_success"].format(input_path=input_path, elapsed_time=f"{(time.time() - start_time):.2f}", output_path=output_path.replace('wav', export_format)))
-
-def change_rms(source_audio, source_rate, target_audio, target_rate, rate):
-    rms2 = F.interpolate(torch.from_numpy(librosa.feature.rms(y=target_audio, frame_length=target_rate // 2 * 2, hop_length=target_rate // 2)).float().unsqueeze(0), size=target_audio.shape[0], mode="linear").squeeze()
-    return (target_audio * (torch.pow(F.interpolate(torch.from_numpy(librosa.feature.rms(y=source_audio, frame_length=source_rate // 2 * 2, hop_length=source_rate // 2)).float().unsqueeze(0), size=target_audio.shape[0], mode="linear").squeeze(), 1 - rate) * torch.pow(torch.maximum(rms2, torch.zeros_like(rms2) + 1e-6), rate - 1)).numpy())
-
-def clear_gpu_cache():
-    gc.collect()
-    if torch.cuda.is_available(): torch.cuda.empty_cache()
-    elif torch.backends.mps.is_available(): torch.mps.empty_cache()
-
-def get_providers():
-    ort_providers = onnxruntime.get_available_providers()
-
-    if "CUDAExecutionProvider" in ort_providers: providers = ["CUDAExecutionProvider"]
-    elif "CoreMLExecutionProvider" in ort_providers: providers = ["CoreMLExecutionProvider"]
-    else: providers = ["CPUExecutionProvider"]
-
-    return providers
-
-class Autotune:
-    def __init__(self, ref_freqs):
-        self.ref_freqs = ref_freqs
-        self.note_dict = self.ref_freqs
-
-    def autotune_f0(self, f0, f0_autotune_strength):
-        autotuned_f0 = np.zeros_like(f0)
-
-        for i, freq in enumerate(f0):
-            autotuned_f0[i] = freq + (min(self.note_dict, key=lambda x: abs(x - freq)) - freq) * f0_autotune_strength
-
-        return autotuned_f0
-
-class VC:
-    def __init__(self, tgt_sr, config):
-        self.x_pad = config.x_pad
-        self.x_query = config.x_query
-        self.x_center = config.x_center
-        self.x_max = config.x_max
-        self.sample_rate = 16000
-        self.window = 160
-        self.t_pad = self.sample_rate * self.x_pad
-        self.t_pad_tgt = tgt_sr * self.x_pad
-        self.t_pad2 = self.t_pad * 2
-        self.t_query = self.sample_rate * self.x_query
-        self.t_center = self.sample_rate * self.x_center
-        self.t_max = self.sample_rate * self.x_max
-        self.time_step = self.window / self.sample_rate * 1000
-        self.f0_min = 50
-        self.f0_max = 1100
-        self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
-        self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
-        self.device = config.device
-        self.is_half = config.is_half
-        self.ref_freqs = [49.00, 51.91, 55.00, 58.27, 61.74, 65.41, 69.30, 73.42, 77.78, 82.41, 87.31, 92.50, 98.00, 103.83, 110.00, 116.54, 123.47, 130.81, 138.59, 146.83, 155.56, 164.81, 174.61, 185.00, 196.00,  207.65, 220.00, 233.08, 246.94, 261.63, 277.18, 293.66, 311.13, 329.63, 349.23, 369.99, 392.00, 415.30, 440.00, 466.16, 493.88, 523.25, 554.37, 587.33, 622.25, 659.25, 698.46, 739.99, 783.99, 830.61, 880.00, 932.33, 987.77, 1046.50]
-        self.autotune = Autotune(self.ref_freqs)
-        self.note_dict = self.autotune.note_dict
-
-    def get_f0_pm(self, x, p_len):
-        import parselmouth
-
-        f0 = (parselmouth.Sound(x, self.sample_rate).to_pitch_ac(time_step=self.window / self.sample_rate * 1000 / 1000, voicing_threshold=0.6, pitch_floor=self.f0_min, pitch_ceiling=self.f0_max).selected_array["frequency"])
-        pad_size = (p_len - len(f0) + 1) // 2
-
-        if pad_size > 0 or p_len - len(f0) - pad_size > 0: f0 = np.pad(f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant")
-        return f0
- 
-    def get_f0_mangio_crepe(self, x, p_len, hop_length, model="full", onnx=False):
-        from main.library.predictors.CREPE import predict
-
-        x = x.astype(np.float32)
-        x /= np.quantile(np.abs(x), 0.999)
-
-        audio = torch.unsqueeze(torch.from_numpy(x).to(self.device, copy=True), dim=0)
-        if audio.ndim == 2 and audio.shape[0] > 1: audio = torch.mean(audio, dim=0, keepdim=True).detach()
-
-        p_len = p_len or x.shape[0] // hop_length
-        source = np.array(predict(audio.detach(), self.sample_rate, hop_length, self.f0_min, self.f0_max, model, batch_size=hop_length * 2, device=self.device, pad=True, providers=get_providers(), onnx=onnx).squeeze(0).cpu().float().numpy())
-        source[source < 0.001] = np.nan
-
-        return np.nan_to_num(np.interp(np.arange(0, len(source) * p_len, len(source)) / p_len, np.arange(0, len(source)), source))
-
-    def get_f0_crepe(self, x, model="full", onnx=False):
-        from main.library.predictors.CREPE import predict, mean, median
-
-        f0, pd = predict(torch.tensor(np.copy(x))[None].float(), self.sample_rate, self.window, self.f0_min, self.f0_max, model, batch_size=512, device=self.device, return_periodicity=True, providers=get_providers(), onnx=onnx)
-        f0, pd = mean(f0, 3), median(pd, 3)
-        f0[pd < 0.1] = 0
-
-        return f0[0].cpu().numpy()
-
-    def get_f0_fcpe(self, x, p_len, hop_length, onnx=False, legacy=False):
-        from main.library.predictors.FCPE import FCPE
-
-        model_fcpe = FCPE(os.path.join("assets", "models", "predictors", ("fcpe_legacy" if legacy else "fcpe") + (".onnx" if onnx else ".pt")), hop_length=int(hop_length), f0_min=int(self.f0_min), f0_max=int(self.f0_max), dtype=torch.float32, device=self.device, sample_rate=self.sample_rate, threshold=0.03 if legacy else 0.006, providers=get_providers(), onnx=onnx, legacy=legacy)
-        f0 = model_fcpe.compute_f0(x, p_len=p_len)
-
-        del model_fcpe
-        return f0
-    
-    def get_f0_rmvpe(self, x, legacy=False, onnx=False):
-        from main.library.predictors.RMVPE import RMVPE
-
-        rmvpe_model = RMVPE(os.path.join("assets", "models", "predictors", "rmvpe" + (".onnx" if onnx else ".pt")), is_half=self.is_half, device=self.device, onnx=onnx, providers=get_providers())
-        f0 = rmvpe_model.infer_from_audio_with_pitch(x, thred=0.03, f0_min=self.f0_min, f0_max=self.f0_max) if legacy else rmvpe_model.infer_from_audio(x, thred=0.03)
-
-        del rmvpe_model
-        return f0
-
-    def get_f0_pyworld(self, x, filter_radius, model="harvest"):
-        from main.library.predictors.WORLD_WRAPPER import PYWORLD
-
-        pw = PYWORLD()
-        x = x.astype(np.double)
-
-        if model == "harvest": f0, t = pw.harvest(x, fs=self.sample_rate, f0_ceil=self.f0_max, f0_floor=self.f0_min, frame_period=10)
-        elif model == "dio": f0, t = pw.dio(x, fs=self.sample_rate, f0_ceil=self.f0_max, f0_floor=self.f0_min, frame_period=10)
-        else: raise ValueError(translations["method_not_valid"])
-
-        f0 = pw.stonemask(x, self.sample_rate, t, f0)
-
-        if filter_radius > 2 or model == "dio": f0 = signal.medfilt(f0, filter_radius)
-        return f0
-    
-    def get_f0_swipe(self, x):
-        from main.library.predictors.SWIPE import swipe
-
-        f0, _ = swipe(x.astype(np.float32), self.sample_rate, f0_floor=self.f0_min, f0_ceil=self.f0_max, frame_period=10)
-        return f0
-    
-    def get_f0_yin(self, x, hop_length, p_len, mode="yin"):
-        source = np.array(librosa.yin(x.astype(np.float32), sr=self.sample_rate, fmin=self.f0_min, fmax=self.f0_max, hop_length=hop_length) if mode == "yin" else librosa.pyin(x.astype(np.float32), fmin=self.f0_min, fmax=self.f0_max, sr=self.sample_rate, hop_length=hop_length)[0])
-        source[source < 0.001] = np.nan
-        return np.nan_to_num(np.interp(np.arange(0, len(source) * p_len, len(source)) / p_len, np.arange(0, len(source)), source))
-
-    def get_f0_hybrid(self, methods_str, x, p_len, hop_length, filter_radius, onnx_mode):
-        methods_str = re.search("hybrid\[(.+)\]", methods_str)
-        if methods_str: methods = [method.strip() for method in methods_str.group(1).split("+")]
-
-        f0_computation_stack, resampled_stack = [], []
-        logger.debug(translations["hybrid_methods"].format(methods=methods))
-
-        x = x.astype(np.float32)
-        x /= np.quantile(np.abs(x), 0.999)
-
-        for method in methods:
-            f0 = None
-            f0_methods = {"pm": lambda: self.get_f0_pm(x, p_len), "dio": lambda: self.get_f0_pyworld(x, filter_radius, "dio"), "mangio-crepe-tiny": lambda: self.get_f0_mangio_crepe(x, p_len, int(hop_length), "tiny", onnx=onnx_mode), "mangio-crepe-small": lambda: self.get_f0_mangio_crepe(x, p_len, int(hop_length), "small", onnx=onnx_mode), "mangio-crepe-medium": lambda: self.get_f0_mangio_crepe(x, p_len, int(hop_length), "medium", onnx=onnx_mode), "mangio-crepe-large": lambda: self.get_f0_mangio_crepe(x, p_len, int(hop_length), "large", onnx=onnx_mode), "mangio-crepe-full": lambda: self.get_f0_mangio_crepe(x, p_len, int(hop_length), "full", onnx=onnx_mode), "crepe-tiny": lambda: self.get_f0_crepe(x, "tiny", onnx=onnx_mode), "crepe-small": lambda: self.get_f0_crepe(x, "small", onnx=onnx_mode), "crepe-medium": lambda: self.get_f0_crepe(x, "medium", onnx=onnx_mode), "crepe-large": lambda: self.get_f0_crepe(x, "large", onnx=onnx_mode), "crepe-full": lambda: self.get_f0_crepe(x, "full", onnx=onnx_mode), "fcpe": lambda: self.get_f0_fcpe(x, p_len, int(hop_length), onnx=onnx_mode), "fcpe-legacy": lambda: self.get_f0_fcpe(x, p_len, int(hop_length), legacy=True, onnx=onnx_mode), "rmvpe": lambda: self.get_f0_rmvpe(x, onnx=onnx_mode), "rmvpe-legacy": lambda: self.get_f0_rmvpe(x, legacy=True, onnx=onnx_mode), "harvest": lambda: self.get_f0_pyworld(x, filter_radius, "harvest"), "yin": lambda: self.get_f0_yin(x, int(hop_length), p_len, mode="yin"), "pyin": lambda: self.get_f0_yin(x, int(hop_length), p_len, mode="pyin"), "swipe": lambda: self.get_f0_swipe(x)}
-            f0 = f0_methods.get(method, lambda: ValueError(translations["method_not_valid"]))()
-            f0_computation_stack.append(f0) 
-
-        for f0 in f0_computation_stack:
-            resampled_stack.append(np.interp(np.linspace(0, len(f0), p_len), np.arange(len(f0)), f0))
-
-        return resampled_stack[0] if len(resampled_stack) == 1 else np.nanmedian(np.vstack(resampled_stack), axis=0)
-
-    def get_f0(self, x, p_len, pitch, f0_method, filter_radius, hop_length, f0_autotune, f0_autotune_strength, inp_f0=None, onnx_mode=False):
-        f0_methods = {"pm": lambda: self.get_f0_pm(x, p_len), "dio": lambda: self.get_f0_pyworld(x, filter_radius, "dio"), "mangio-crepe-tiny": lambda: self.get_f0_mangio_crepe(x, p_len, int(hop_length), "tiny", onnx=onnx_mode), "mangio-crepe-small": lambda: self.get_f0_mangio_crepe(x, p_len, int(hop_length), "small", onnx=onnx_mode), "mangio-crepe-medium": lambda: self.get_f0_mangio_crepe(x, p_len, int(hop_length), "medium", onnx=onnx_mode), "mangio-crepe-large": lambda: self.get_f0_mangio_crepe(x, p_len, int(hop_length), "large", onnx=onnx_mode), "mangio-crepe-full": lambda: self.get_f0_mangio_crepe(x, p_len, int(hop_length), "full", onnx=onnx_mode), "crepe-tiny": lambda: self.get_f0_crepe(x, "tiny", onnx=onnx_mode), "crepe-small": lambda: self.get_f0_crepe(x, "small", onnx=onnx_mode), "crepe-medium": lambda: self.get_f0_crepe(x, "medium", onnx=onnx_mode), "crepe-large": lambda: self.get_f0_crepe(x, "large", onnx=onnx_mode), "crepe-full": lambda: self.get_f0_crepe(x, "full", onnx=onnx_mode), "fcpe": lambda: self.get_f0_fcpe(x, p_len, int(hop_length), onnx=onnx_mode), "fcpe-legacy": lambda: self.get_f0_fcpe(x, p_len, int(hop_length), legacy=True, onnx=onnx_mode), "rmvpe": lambda: self.get_f0_rmvpe(x, onnx=onnx_mode), "rmvpe-legacy": lambda: self.get_f0_rmvpe(x, legacy=True, onnx=onnx_mode), "harvest": lambda: self.get_f0_pyworld(x, filter_radius, "harvest"), "yin": lambda: self.get_f0_yin(x, int(hop_length), p_len, mode="yin"), "pyin": lambda: self.get_f0_yin(x, int(hop_length), p_len, mode="pyin"), "swipe": lambda: self.get_f0_swipe(x)}
-        f0 = self.get_f0_hybrid(f0_method, x, p_len, hop_length, filter_radius, onnx_mode) if "hybrid" in f0_method else f0_methods.get(f0_method, lambda: ValueError(translations["method_not_valid"]))()
-
-        if f0_autotune: f0 = Autotune.autotune_f0(self, f0, f0_autotune_strength)
-        if isinstance(f0, tuple): f0 = f0[0]
-
-        f0 *= pow(2, pitch / 12)
-        tf0 = self.sample_rate // self.window
-
-        if inp_f0 is not None:
-            replace_f0 = np.interp(list(range(np.round((inp_f0[:, 0].max() - inp_f0[:, 0].min()) * tf0 + 1).astype(np.int16))), inp_f0[:, 0] * 100, inp_f0[:, 1])
-            f0[self.x_pad * tf0 : self.x_pad * tf0 + len(replace_f0)] = replace_f0[:f0[self.x_pad * tf0 : self.x_pad * tf0 + len(replace_f0)].shape[0]]
-
-        f0_mel = 1127 * np.log(1 + f0 / 700)
-        f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - self.f0_mel_min) * 254 / (self.f0_mel_max - self.f0_mel_min) + 1
-        f0_mel[f0_mel <= 1] = 1
-        f0_mel[f0_mel > 255] = 255
-
-        return np.rint(f0_mel).astype(np.int32), f0.copy()
-    
-    def extract_features(self, model, feats, version):
-        return torch.as_tensor(model.run([model.get_outputs()[0].name, model.get_outputs()[1].name], {"feats": feats.detach().cpu().numpy()})[0 if version == "v1" else 1], dtype=torch.float32, device=feats.device)
-
-    def voice_conversion(self, model, net_g, sid, audio0, pitch, pitchf, index, big_npy, index_rate, version, protect):
-        pitch_guidance = pitch != None and pitchf != None
-        feats = (torch.from_numpy(audio0).half() if self.is_half else torch.from_numpy(audio0).float())
-
-        if feats.dim() == 2: feats = feats.mean(-1)
-        assert feats.dim() == 1, feats.dim()
-        feats = feats.view(1, -1)
-
-        with torch.no_grad():
-            if self.embed_suffix == ".pt":
-                padding_mask = torch.BoolTensor(feats.shape).to(self.device).fill_(False)
-                logits = model.extract_features(**{"source": feats.to(self.device), "padding_mask": padding_mask, "output_layer": 9 if version == "v1" else 12})
-                feats = model.final_proj(logits[0]) if version == "v1" else logits[0]
-            elif self.embed_suffix == ".onnx": feats = self.extract_features(model, feats.to(self.device), version).to(self.device)
-            elif self.embed_suffix == ".safetensors":
-                logits = model(feats.to(self.device))["last_hidden_state"]
-                feats = (model.final_proj(logits[0]).unsqueeze(0) if version == "v1" else logits)
-            else: raise ValueError(translations["option_not_valid"])
-
-            if protect < 0.5 and pitch_guidance: feats0 = feats.clone()
-
-            if (not isinstance(index, type(None)) and not isinstance(big_npy, type(None)) and index_rate != 0):
-                npy = feats[0].cpu().numpy()
-                if self.is_half: npy = npy.astype(np.float32)
-
-                score, ix = index.search(npy, k=8)
-                weight = np.square(1 / score)
-
-                npy = np.sum(big_npy[ix] * np.expand_dims(weight / weight.sum(axis=1, keepdims=True), axis=2), axis=1)
-                if self.is_half: npy = npy.astype(np.float16)
-
-                feats = (torch.from_numpy(npy).unsqueeze(0).to(self.device) * index_rate + (1 - index_rate) * feats)
-
-            feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
-            if protect < 0.5 and pitch_guidance: feats0 = F.interpolate(feats0.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
-
-            p_len = audio0.shape[0] // self.window
-            
-            if feats.shape[1] < p_len:
-                p_len = feats.shape[1]
-                if pitch_guidance:
-                    pitch = pitch[:, :p_len]
-                    pitchf = pitchf[:, :p_len]
-
-            if protect < 0.5 and pitch_guidance:
-                pitchff = pitchf.clone()
-                pitchff[pitchf > 0] = 1
-                pitchff[pitchf < 1] = protect
-                pitchff = pitchff.unsqueeze(-1)
-
-                feats = (feats * pitchff + feats0 * (1 - pitchff)).to(feats0.dtype)
-
-            p_len = torch.tensor([p_len], device=self.device).long()
-            audio1 = ((net_g.infer(feats.half() if self.is_half else feats.float(), p_len, pitch if pitch_guidance else None, (pitchf.half() if self.is_half else pitchf.float()) if pitch_guidance else None, sid)[0][0, 0]).data.cpu().float().numpy()) if self.suffix == ".pth" else (net_g.run([net_g.get_outputs()[0].name], ({net_g.get_inputs()[0].name: feats.cpu().numpy().astype(np.float32), net_g.get_inputs()[1].name: p_len.cpu().numpy(), net_g.get_inputs()[2].name: np.array([sid.cpu().item()], dtype=np.int64), net_g.get_inputs()[3].name: np.random.randn(1, 192, p_len).astype(np.float32), net_g.get_inputs()[4].name: pitch.cpu().numpy().astype(np.int64), net_g.get_inputs()[5].name: pitchf.cpu().numpy().astype(np.float32)} if pitch_guidance else {net_g.get_inputs()[0].name: feats.cpu().numpy().astype(np.float32), net_g.get_inputs()[1].name: p_len.cpu().numpy(), net_g.get_inputs()[2].name: np.array([sid.cpu().item()], dtype=np.int64), net_g.get_inputs()[3].name: np.random.randn(1, 192, p_len).astype(np.float32)}))[0][0, 0])
-
-        if self.embed_suffix == ".pt": del padding_mask
-        del feats, p_len, net_g
-        clear_gpu_cache()
-        return audio1
-    
-    def pipeline(self, model, net_g, sid, audio, pitch, f0_method, file_index, index_rate, pitch_guidance, filter_radius, volume_envelope, version, protect, hop_length, f0_autotune, f0_autotune_strength, suffix, embed_suffix, f0_file=None, f0_onnx=False, pbar=None):
-        self.suffix = suffix
-        self.embed_suffix = embed_suffix
-
-        if file_index != "" and os.path.exists(file_index) and index_rate != 0:
-            try:
-                index = faiss.read_index(file_index)
-                big_npy = index.reconstruct_n(0, index.ntotal)
-            except Exception as e:
-                logger.error(translations["read_faiss_index_error"].format(e=e))
-                index = big_npy = None
-        else: index = big_npy = None
-
-        pbar.update(1)
-        opt_ts, audio_opt = [], []
-        audio = signal.filtfilt(bh, ah, audio)
-        audio_pad = np.pad(audio, (self.window // 2, self.window // 2), mode="reflect")
-
-        if audio_pad.shape[0] > self.t_max:
-            audio_sum = np.zeros_like(audio)
-            for i in range(self.window):
-                audio_sum += audio_pad[i : i - self.window]
-
-            for t in range(self.t_center, audio.shape[0], self.t_center):
-                opt_ts.append(t - self.t_query + np.where(np.abs(audio_sum[t - self.t_query : t + self.t_query]) == np.abs(audio_sum[t - self.t_query : t + self.t_query]).min())[0][0])
-
-        s = 0
-        t, inp_f0 = None, None
-        audio_pad = np.pad(audio, (self.t_pad, self.t_pad), mode="reflect")
-        sid = torch.tensor(sid, device=self.device).unsqueeze(0).long()
-        p_len = audio_pad.shape[0] // self.window
-
-        if hasattr(f0_file, "name"):
-            try:
-                with open(f0_file.name, "r") as f:
-                    raw_lines = f.read()
-                    if len(raw_lines) > 0:
-                        inp_f0 = []
-                        for line in raw_lines.strip("\n").split("\n"):
-                            inp_f0.append([float(i) for i in line.split(",")])
-
-                        inp_f0 = np.array(inp_f0, dtype=np.float32)
-            except:
-                logger.error(translations["error_readfile"])
-                inp_f0 = None
-
-        pbar.update(1)
-        if pitch_guidance:
-            pitch, pitchf = self.get_f0(audio_pad, p_len, pitch, f0_method, filter_radius, hop_length, f0_autotune, f0_autotune_strength, inp_f0, onnx_mode=f0_onnx)
-            pitch, pitchf = pitch[:p_len], pitchf[:p_len]
-            if self.device == "mps": pitchf = pitchf.astype(np.float32)
-            pitch, pitchf = torch.tensor(pitch, device=self.device).unsqueeze(0).long(), torch.tensor(pitchf, device=self.device).unsqueeze(0).float()
-
-        pbar.update(1)
-        for t in opt_ts:
-            t = t // self.window * self.window
-            audio_opt.append(self.voice_conversion(model, net_g, sid, audio_pad[s : t + self.t_pad2 + self.window], pitch[:, s // self.window : (t + self.t_pad2) // self.window] if pitch_guidance else None, pitchf[:, s // self.window : (t + self.t_pad2) // self.window] if pitch_guidance else None, index, big_npy, index_rate, version, protect)[self.t_pad_tgt : -self.t_pad_tgt])    
-            s = t
-            
-        audio_opt.append(self.voice_conversion(model, net_g, sid, audio_pad[t:], (pitch[:, t // self.window :] if t is not None else pitch) if pitch_guidance else None, (pitchf[:, t // self.window :] if t is not None else pitchf) if pitch_guidance else None, index, big_npy, index_rate, version, protect)[self.t_pad_tgt : -self.t_pad_tgt])
-        audio_opt = np.concatenate(audio_opt)
-        if volume_envelope != 1: audio_opt = change_rms(audio, self.sample_rate, audio_opt, self.sample_rate, volume_envelope)
-        audio_max = np.abs(audio_opt).max() / 0.99
-        if audio_max > 1: audio_opt /= audio_max
-
-        if pitch_guidance: del pitch, pitchf
-        del sid
-        clear_gpu_cache()
-        pbar.update(1)
-
-        return audio_opt
-
-class VoiceConverter:
-    def __init__(self, model_path, sid = 0):
-        self.config = config
-        self.device = config.device
-        self.hubert_model = None
-        self.tgt_sr = None 
-        self.net_g = None 
-        self.vc = None
-        self.cpt = None  
-        self.version = None 
-        self.n_spk = None  
-        self.use_f0 = None  
-        self.loaded_model = None
-        self.vocoder = "Default"
-        self.checkpointing = False
-        self.sample_rate = 16000
-        self.sid = sid
-        self.get_vc(model_path, sid)
-
-    def convert_audio(self, audio_input_path, audio_output_path, index_path, embedder_model, pitch, f0_method, index_rate, volume_envelope, protect, hop_length, f0_autotune, f0_autotune_strength, filter_radius, clean_audio, clean_strength, export_format, resample_sr = 0, checkpointing = False, f0_file = None, f0_onnx = False, embedders_mode = "fairseq", formant_shifting = False, formant_qfrency = 0.8, formant_timbre = 0.8, split_audio = False):
-        try:
-            with tqdm(total=10, desc=translations["convert_audio"], ncols=100, unit="a") as pbar:
-                audio = load_audio(logger, audio_input_path, self.sample_rate, formant_shifting=formant_shifting, formant_qfrency=formant_qfrency, formant_timbre=formant_timbre)
-                self.checkpointing = checkpointing
-                audio_max = np.abs(audio).max() / 0.95
-                if audio_max > 1: audio /= audio_max
-
-                pbar.update(1)
-                if not self.hubert_model:
-                    models, _, embed_suffix = load_embedders_model(embedder_model, embedders_mode, providers=get_providers())
-                    self.hubert_model = (models.to(self.device).half() if self.config.is_half else models.to(self.device).float()).eval() if embed_suffix in [".pt", ".safetensors"] else models
-                    self.embed_suffix = embed_suffix
-
-                pbar.update(1)
-                if self.tgt_sr != resample_sr >= self.sample_rate: self.tgt_sr = resample_sr
-                target_sr = min([8000, 11025, 12000, 16000, 22050, 24000, 32000, 44100, 48000, 96000], key=lambda x: abs(x - self.tgt_sr))
-
-                if split_audio:
-                    chunks = cut(audio, self.sample_rate, db_thresh=-60, min_interval=500)  
-                    pbar.total = len(chunks) * 4 + 6
-                    logger.info(f"{translations['split_total']}: {len(chunks)}")
-                else: chunks = [(audio, 0, 0)]
-
-                converted_chunks = []
-                pbar.update(1)
-
-                for waveform, start, end in chunks:
-                    converted_chunks.append((start, end, self.vc.pipeline(model=self.hubert_model, net_g=self.net_g, sid=self.sid, audio=waveform, pitch=pitch, f0_method=f0_method, file_index=(index_path.strip().strip('"').strip("\n").strip('"').strip().replace("trained", "added")), index_rate=index_rate, pitch_guidance=self.use_f0, filter_radius=filter_radius, volume_envelope=volume_envelope, version=self.version, protect=protect, hop_length=hop_length, f0_autotune=f0_autotune, f0_autotune_strength=f0_autotune_strength, suffix=self.suffix, embed_suffix=self.embed_suffix, f0_file=f0_file, f0_onnx=f0_onnx, pbar=pbar)))
-                
-                pbar.update(1)
-                audio_output = restore(converted_chunks, total_len=len(audio), dtype=converted_chunks[0][2].dtype) if split_audio else converted_chunks[0][2]
-                if target_sr >= self.sample_rate and self.tgt_sr != target_sr: audio_output = librosa.resample(audio_output, orig_sr=self.tgt_sr, target_sr=target_sr, res_type="soxr_vhq")
-
-                pbar.update(1)
-                if clean_audio:
-                    from main.tools.noisereduce import reduce_noise
-                    audio_output = reduce_noise(y=audio_output, sr=target_sr, prop_decrease=clean_strength, device=self.device) 
-
-                sf.write(audio_output_path, audio_output, target_sr, format=export_format)
-                pbar.update(1)
-        except Exception as e:
-            logger.error(translations["error_convert"].format(e=e))
-            import traceback
-            logger.debug(traceback.format_exc())
-
-    def get_vc(self, weight_root, sid):
-        if sid == "" or sid == []:
-            self.cleanup()
-            clear_gpu_cache()
-
-        if not self.loaded_model or self.loaded_model != weight_root:
-            self.loaded_model = weight_root
-            self.load_model()
-            if self.cpt is not None: self.setup()
-
-    def cleanup(self):
-        if self.hubert_model is not None:
-            del self.net_g, self.n_spk, self.vc, self.hubert_model, self.tgt_sr
-            self.hubert_model = self.net_g = self.n_spk = self.vc = self.tgt_sr = None
-            clear_gpu_cache()
-
-        del self.net_g, self.cpt
-        clear_gpu_cache()
-        self.cpt = None
-
-    def load_model(self):
-        if os.path.isfile(self.loaded_model):
-            if self.loaded_model.endswith(".pth"): self.cpt = torch.load(self.loaded_model, map_location="cpu")  
-            else: 
-                sess_options = onnxruntime.SessionOptions()
-                sess_options.log_severity_level = 3
-                self.cpt = onnxruntime.InferenceSession(self.loaded_model, sess_options=sess_options, providers=get_providers())
-        else: self.cpt = None
-
-    def setup(self):
-        if self.cpt is not None:
-            if self.loaded_model.endswith(".pth"):
-                self.tgt_sr = self.cpt["config"][-1]
-                self.cpt["config"][-3] = self.cpt["weight"]["emb_g.weight"].shape[0]
-                self.use_f0 = self.cpt.get("f0", 1)
-                self.version = self.cpt.get("version", "v1")
-                self.vocoder = self.cpt.get("vocoder", "Default")
-                if self.vocoder != "Default": self.config.is_half = False
-
-                self.net_g = Synthesizer(*self.cpt["config"], use_f0=self.use_f0, text_enc_hidden_dim=768 if self.version == "v2" else 256, vocoder=self.vocoder, checkpointing=self.checkpointing)
-                del self.net_g.enc_q
-
-                self.net_g.load_state_dict(self.cpt["weight"], strict=False)
-                self.net_g.eval().to(self.device)
-                self.net_g = (self.net_g.half() if self.config.is_half else self.net_g.float())
-                self.n_spk = self.cpt["config"][-3]
-                self.suffix = ".pth"
-            else:
-                import json
-                import onnx
-
-                metadata_dict = None
-                for prop in onnx.load(self.loaded_model).metadata_props:
-                    if prop.key == "model_info":
-                        metadata_dict = json.loads(prop.value)
-                        break
-
-                self.net_g = self.cpt
-                self.tgt_sr = metadata_dict.get("sr", 32000)
-                self.use_f0 = metadata_dict.get("f0", 1)
-                self.version = metadata_dict.get("version", "v1")
-                self.suffix = ".onnx"
-
-            self.vc = VC(self.tgt_sr, self.config)
-
-if __name__ == "__main__": main()

main/inference/create_dataset.py

@@ -1,230 +0,0 @@
-import os
-import sys
-import time
-import yt_dlp
-import shutil
-import librosa
-import logging
-import argparse
-import warnings
-import logging.handlers
-
-from soundfile import read, write
-from distutils.util import strtobool
-
-sys.path.append(os.getcwd())
-
-from main.configs.config import Config
-from main.library.algorithm.separator import Separator
-
-config = Config()
-translations = config.translations
-dataset_temp = os.path.join("dataset_temp")
-logger = logging.getLogger(__name__)
-
-if logger.hasHandlers(): logger.handlers.clear()
-else: 
-    console_handler = logging.StreamHandler()
-    console_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
-    console_handler.setFormatter(console_formatter)
-    console_handler.setLevel(logging.INFO)
-    file_handler = logging.handlers.RotatingFileHandler(os.path.join("assets", "logs", "create_dataset.log"), maxBytes=5*1024*1024, backupCount=3, encoding='utf-8')
-    file_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
-    file_handler.setFormatter(file_formatter)
-    file_handler.setLevel(logging.DEBUG)
-    logger.addHandler(console_handler)
-    logger.addHandler(file_handler)
-    logger.setLevel(logging.DEBUG)
-
-def parse_arguments():
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--input_audio", type=str, required=True)
-    parser.add_argument("--output_dataset", type=str, default="./dataset")
-    parser.add_argument("--sample_rate", type=int, default=44100)
-    parser.add_argument("--clean_dataset", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--clean_strength", type=float, default=0.7)
-    parser.add_argument("--separator_reverb", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--kim_vocal_version", type=int, default=2)
-    parser.add_argument("--overlap", type=float, default=0.25)
-    parser.add_argument("--segments_size", type=int, default=256)
-    parser.add_argument("--mdx_hop_length", type=int, default=1024)
-    parser.add_argument("--mdx_batch_size", type=int, default=1)
-    parser.add_argument("--denoise_mdx", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--skip", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--skip_start_audios", type=str, default="0")
-    parser.add_argument("--skip_end_audios", type=str, default="0")
-
-    return parser.parse_args()
-
-def main():
-    pid_path = os.path.join("assets", "create_dataset_pid.txt")
-    with open(pid_path, "w") as pid_file:
-        pid_file.write(str(os.getpid()))
-
-    args = parse_arguments()
-    input_audio, output_dataset, sample_rate, clean_dataset, clean_strength, separator_reverb, kim_vocal_version, overlap, segments_size, hop_length, batch_size, denoise_mdx, skip, skip_start_audios, skip_end_audios = args.input_audio, args.output_dataset, args.sample_rate, args.clean_dataset, args.clean_strength, args.separator_reverb, args.kim_vocal_version, args.overlap, args.segments_size, args.mdx_hop_length, args.mdx_batch_size, args.denoise_mdx, args.skip, args.skip_start_audios, args.skip_end_audios
-    log_data = {translations['audio_path']: input_audio, translations['output_path']: output_dataset, translations['sr']: sample_rate, translations['clear_dataset']: clean_dataset, translations['dereveb_audio']: separator_reverb, translations['segments_size']: segments_size, translations['overlap']: overlap, "Hop length": hop_length, translations['batch_size']: batch_size, translations['denoise_mdx']: denoise_mdx, translations['skip']: skip}
-
-    if clean_dataset: log_data[translations['clean_strength']] = clean_strength
-    if skip:
-        log_data[translations['skip_start']] = skip_start_audios
-        log_data[translations['skip_end']] = skip_end_audios
-
-    for key, value in log_data.items():
-        logger.debug(f"{key}: {value}")
-
-    if kim_vocal_version not in [1, 2]: raise ValueError(translations["version_not_valid"])
-    start_time = time.time()
-
-    try:
-        paths = []
-
-        if not os.path.exists(dataset_temp): os.makedirs(dataset_temp, exist_ok=True)
-        urls = input_audio.replace(", ", ",").split(",")
-
-        for url in urls:
-            path = downloader(url, urls.index(url))
-            paths.append(path)
-
-        if skip:
-            skip_start_audios, skip_end_audios = skip_start_audios.replace(", ", ",").split(","), skip_end_audios.replace(", ", ",").split(",")
-
-            if len(skip_start_audios) < len(paths) or len(skip_end_audios) < len(paths): 
-                logger.warning(translations["skip<audio"])
-                sys.exit(1)
-            elif len(skip_start_audios) > len(paths) or len(skip_end_audios) > len(paths): 
-                logger.warning(translations["skip>audio"])
-                sys.exit(1)
-            else:
-                for audio, skip_start_audio, skip_end_audio in zip(paths, skip_start_audios, skip_end_audios):
-                    skip_start(audio, skip_start_audio)
-                    skip_end(audio, skip_end_audio)
-
-        separator_paths = []
-
-        for audio in paths:
-            vocals = separator_music_main(audio, dataset_temp, segments_size, overlap, denoise_mdx, kim_vocal_version, hop_length, batch_size, sample_rate)
-            if separator_reverb: vocals = separator_reverb_audio(vocals, dataset_temp, segments_size, overlap, denoise_mdx, hop_length, batch_size, sample_rate)
-            separator_paths.append(vocals)
-        
-        paths = separator_paths
-
-        for audio_path in paths:
-            data, sample_rate = read(audio_path)
-            data = librosa.to_mono(data.T)
-            
-            if clean_dataset: 
-                from main.tools.noisereduce import reduce_noise
-                data = reduce_noise(y=data, sr=sample_rate, prop_decrease=clean_strength, device=config.device)
-
-            write(audio_path, data, sample_rate)
-    except Exception as e:
-        logger.error(f"{translations['create_dataset_error']}: {e}")
-        import traceback
-        logger.error(traceback.format_exc())
-    finally:
-        for audio in paths:
-            shutil.move(audio, output_dataset)
-
-        if os.path.exists(dataset_temp): shutil.rmtree(dataset_temp, ignore_errors=True)
-
-    elapsed_time = time.time() - start_time
-    if os.path.exists(pid_path): os.remove(pid_path)
-    logger.info(translations["create_dataset_success"].format(elapsed_time=f"{elapsed_time:.2f}"))
-
-def downloader(url, name):
-    with warnings.catch_warnings():
-        warnings.simplefilter("ignore")
-
-        ydl_opts = {"format": "bestaudio/best", "outtmpl": os.path.join(dataset_temp, f"{name}"), "postprocessors": [{"key": "FFmpegExtractAudio", "preferredcodec": "wav", "preferredquality": "192"}], "no_warnings": True, "noplaylist": True, "noplaylist": True, "verbose": False}
-        logger.info(f"{translations['starting_download']}: {url}...")
-
-        with yt_dlp.YoutubeDL(ydl_opts) as ydl:
-            ydl.extract_info(url)  
-            logger.info(f"{translations['download_success']}: {url}")
-
-    return os.path.join(dataset_temp, f"{name}" + ".wav")
-
-def skip_start(input_file, seconds):
-    data, sr = read(input_file)
-    total_duration = len(data) / sr
-    
-    if seconds <= 0: logger.warning(translations["=<0"])
-    elif seconds >= total_duration: logger.warning(translations["skip_warning"].format(seconds=seconds, total_duration=f"{total_duration:.2f}"))
-    else: 
-        logger.info(f"{translations['skip_start']}: {input_file}...")
-        write(input_file, data[int(seconds * sr):], sr)
-
-        logger.info(translations["skip_start_audio"].format(input_file=input_file))
-
-def skip_end(input_file, seconds):
-    data, sr = read(input_file)
-    total_duration = len(data) / sr
-
-    if seconds <= 0: logger.warning(translations["=<0"])
-    elif seconds > total_duration: logger.warning(translations["skip_warning"].format(seconds=seconds, total_duration=f"{total_duration:.2f}"))
-    else: 
-        logger.info(f"{translations['skip_end']}: {input_file}...")
-        write(input_file, data[:-int(seconds * sr)], sr)
-
-        logger.info(translations["skip_end_audio"].format(input_file=input_file))
-
-def separator_music_main(input, output, segments_size, overlap, denoise, version, hop_length, batch_size, sample_rate):
-    if not os.path.exists(input): 
-        logger.warning(translations["input_not_valid"])
-        return None
-    
-    if not os.path.exists(output): 
-        logger.warning(translations["output_not_valid"])
-        return None
-
-    model = f"Kim_Vocal_{version}.onnx"
-    output_separator = separator_main(audio_file=input, model_filename=model, output_format="wav", output_dir=output, mdx_segment_size=segments_size, mdx_overlap=overlap, mdx_batch_size=batch_size, mdx_hop_length=hop_length, mdx_enable_denoise=denoise, sample_rate=sample_rate)
-
-    for f in output_separator:
-        path = os.path.join(output, f)
-        if not os.path.exists(path): logger.error(translations["not_found"].format(name=path))
-
-        if '_(Instrumental)_' in f: os.rename(path, os.path.splitext(path)[0].replace("(", "").replace(")", "") + ".wav")
-        elif '_(Vocals)_' in f:
-            rename_file = os.path.splitext(path)[0].replace("(", "").replace(")", "") + ".wav"
-            os.rename(path, rename_file)
-
-    return rename_file
-
-def separator_reverb_audio(input, output, segments_size, overlap, denoise, hop_length, batch_size, sample_rate):
-    if not os.path.exists(input): 
-        logger.warning(translations["input_not_valid"])
-        return None
-    
-    if not os.path.exists(output): 
-        logger.warning(translations["output_not_valid"])
-        return None
-
-    logger.info(f"{translations['dereverb']}: {input}...")
-    output_dereverb = separator_main(audio_file=input, model_filename="Reverb_HQ_By_FoxJoy.onnx", output_format="wav", output_dir=output, mdx_segment_size=segments_size, mdx_overlap=overlap, mdx_batch_size=hop_length, mdx_hop_length=batch_size, mdx_enable_denoise=denoise, sample_rate=sample_rate)
-
-    for f in output_dereverb:
-        path = os.path.join(output, f)
-        if not os.path.exists(path): logger.error(translations["not_found"].format(name=path))
-
-        if '_(Reverb)_' in f: os.rename(path, os.path.splitext(path)[0].replace("(", "").replace(")", "") + ".wav")
-        elif '_(No Reverb)_' in f:
-            rename_file = os.path.splitext(path)[0].replace("(", "").replace(")", "") + ".wav"
-            os.rename(path, rename_file)    
-
-    logger.info(f"{translations['dereverb_success']}: {rename_file}")
-    return rename_file
-
-def separator_main(audio_file=None, model_filename="Kim_Vocal_1.onnx", output_format="wav", output_dir=".", mdx_segment_size=256, mdx_overlap=0.25, mdx_batch_size=1, mdx_hop_length=1024, mdx_enable_denoise=True, sample_rate=44100):
-    try:
-        separator = Separator(logger=logger, log_formatter=file_formatter, log_level=logging.INFO, output_dir=output_dir, output_format=output_format, output_bitrate=None, normalization_threshold=0.9, output_single_stem=None, invert_using_spec=False, sample_rate=sample_rate, mdx_params={"hop_length": mdx_hop_length, "segment_size": mdx_segment_size, "overlap": mdx_overlap, "batch_size": mdx_batch_size, "enable_denoise": mdx_enable_denoise})
-        separator.load_model(model_filename=model_filename)
-        return separator.separate(audio_file)
-    except:
-        logger.debug(translations["default_setting"])
-        separator = Separator(logger=logger, log_formatter=file_formatter, log_level=logging.INFO, output_dir=output_dir, output_format=output_format, output_bitrate=None, normalization_threshold=0.9, output_single_stem=None, invert_using_spec=False, sample_rate=44100, mdx_params={"hop_length": 1024, "segment_size": 256, "overlap": 0.25, "batch_size": 1, "enable_denoise": mdx_enable_denoise})
-        separator.load_model(model_filename=model_filename)
-        return separator.separate(audio_file)
-
-if __name__ == "__main__": main()

main/inference/create_index.py

@@ -1,90 +0,0 @@
-import os
-import sys
-import faiss
-import logging
-import argparse
-import logging.handlers
-
-import numpy as np
-
-from multiprocessing import cpu_count
-from sklearn.cluster import MiniBatchKMeans
-
-sys.path.append(os.getcwd())
-
-from main.configs.config import Config
-translations = Config().translations
-
-def parse_arguments():
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--model_name", type=str, required=True)
-    parser.add_argument("--rvc_version", type=str, default="v2")
-    parser.add_argument("--index_algorithm", type=str, default="Auto")
-
-    return parser.parse_args()
-
-def main():
-    args = parse_arguments()
-    exp_dir = os.path.join("assets", "logs", args.model_name)
-    version, index_algorithm = args.rvc_version, args.index_algorithm
-    logger = logging.getLogger(__name__)
-
-    if logger.hasHandlers(): logger.handlers.clear()
-    else:  
-        console_handler = logging.StreamHandler()
-        console_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
-        console_handler.setFormatter(console_formatter)
-        console_handler.setLevel(logging.INFO)
-        file_handler = logging.handlers.RotatingFileHandler(os.path.join(exp_dir, "create_index.log"), maxBytes=5*1024*1024, backupCount=3, encoding='utf-8')
-        file_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
-        file_handler.setFormatter(file_formatter)
-        file_handler.setLevel(logging.DEBUG)
-        logger.addHandler(console_handler)
-        logger.addHandler(file_handler)
-        logger.setLevel(logging.DEBUG)
-
-    log_data = {translations['modelname']: args.model_name, translations['model_path']: exp_dir, translations['training_version']: version, translations['index_algorithm_info']: index_algorithm}
-    for key, value in log_data.items():
-        logger.debug(f"{key}: {value}")
-
-    try:
-        npys = []
-        feature_dir = os.path.join(exp_dir, f"{version}_extracted")
-        model_name = os.path.basename(exp_dir)
-
-        for name in sorted(os.listdir(feature_dir)):
-            npys.append(np.load(os.path.join(feature_dir, name)))
-
-        big_npy = np.concatenate(npys, axis=0)
-        big_npy_idx = np.arange(big_npy.shape[0])
-        np.random.shuffle(big_npy_idx)
-        big_npy = big_npy[big_npy_idx]
-
-        if big_npy.shape[0] > 2e5 and (index_algorithm == "Auto" or index_algorithm == "KMeans"): big_npy = (MiniBatchKMeans(n_clusters=10000, verbose=True, batch_size=256 * cpu_count(), compute_labels=False, init="random").fit(big_npy).cluster_centers_)
-        np.save(os.path.join(exp_dir, "total_fea.npy"), big_npy)
-
-        n_ivf = min(int(16 * np.sqrt(big_npy.shape[0])), big_npy.shape[0] // 39)
-        index_trained = faiss.index_factory(256 if version == "v1" else 768, f"IVF{n_ivf},Flat")
-        index_ivf_trained = faiss.extract_index_ivf(index_trained)
-        index_ivf_trained.nprobe = 1
-        index_trained.train(big_npy)
-        faiss.write_index(index_trained, os.path.join(exp_dir, f"trained_IVF{n_ivf}_Flat_nprobe_{index_ivf_trained.nprobe}_{model_name}_{version}.index"))
-
-        index_added = faiss.index_factory(256 if version == "v1" else 768, f"IVF{n_ivf},Flat")
-        index_ivf_added = faiss.extract_index_ivf(index_added)
-        index_ivf_added.nprobe = 1
-        index_added.train(big_npy)
-        batch_size_add = 8192
-    
-        for i in range(0, big_npy.shape[0], batch_size_add):
-            index_added.add(big_npy[i : i + batch_size_add])
-
-        index_filepath_added = os.path.join(exp_dir, f"added_IVF{n_ivf}_Flat_nprobe_{index_ivf_added.nprobe}_{model_name}_{version}.index")
-        faiss.write_index(index_added, index_filepath_added)
-        logger.info(f"{translations['save_index']} '{index_filepath_added}'")
-    except Exception as e:
-        logger.error(f"{translations['create_index_error']}: {e}")
-        import traceback
-        logger.debug(traceback.format_exc())
-
-if __name__ == "__main__": main()

main/inference/extract.py

@@ -1,360 +0,0 @@
-import os
-import re
-import sys
-import time
-import tqdm
-import torch
-import shutil
-import logging
-import argparse
-import warnings
-import onnxruntime
-import logging.handlers
-
-import numpy as np
-import soundfile as sf
-import torch.nn.functional as F
-
-from random import shuffle
-from distutils.util import strtobool
-from concurrent.futures import ThreadPoolExecutor, as_completed
-
-sys.path.append(os.getcwd())
-
-from main.configs.config import Config
-from main.library.utils import check_predictors, check_embedders, load_audio, load_embedders_model
-
-logger = logging.getLogger(__name__)
-config = Config()
-translations = config.translations
-logger.propagate = False
-
-warnings.filterwarnings("ignore")
-for l in ["torch", "faiss", "httpx", "httpcore", "faiss.loader", "numba.core", "urllib3", "matplotlib"]:
-    logging.getLogger(l).setLevel(logging.ERROR)
-
-def parse_arguments():
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--model_name", type=str, required=True)
-    parser.add_argument("--rvc_version", type=str, default="v2")
-    parser.add_argument("--f0_method", type=str, default="rmvpe")
-    parser.add_argument("--pitch_guidance", type=lambda x: bool(strtobool(x)), default=True)
-    parser.add_argument("--hop_length", type=int, default=128)
-    parser.add_argument("--cpu_cores", type=int, default=2)
-    parser.add_argument("--gpu", type=str, default="-")
-    parser.add_argument("--sample_rate", type=int, required=True)
-    parser.add_argument("--embedder_model", type=str, default="contentvec_base")
-    parser.add_argument("--f0_onnx", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--embedders_mode", type=str, default="fairseq")
-
-    return parser.parse_args()
-
-def generate_config(rvc_version, sample_rate, model_path):
-    config_save_path = os.path.join(model_path, "config.json")
-    if not os.path.exists(config_save_path): shutil.copy(os.path.join("main", "configs", rvc_version, f"{sample_rate}.json"), config_save_path)
-
-def generate_filelist(pitch_guidance, model_path, rvc_version, sample_rate, embedders_mode = "fairseq"):
-    gt_wavs_dir, feature_dir = os.path.join(model_path, "sliced_audios"), os.path.join(model_path, f"{rvc_version}_extracted")
-    f0_dir, f0nsf_dir = None, None
-    
-    if pitch_guidance: f0_dir, f0nsf_dir = os.path.join(model_path, "f0"), os.path.join(model_path, "f0_voiced")
-
-    gt_wavs_files, feature_files = set(name.split(".")[0] for name in os.listdir(gt_wavs_dir)), set(name.split(".")[0] for name in os.listdir(feature_dir))
-    names = gt_wavs_files & feature_files & set(name.split(".")[0] for name in os.listdir(f0_dir)) & set(name.split(".")[0] for name in os.listdir(f0nsf_dir)) if pitch_guidance else gt_wavs_files & feature_files
-
-    options = []
-    mute_base_path = os.path.join("assets", "logs", "mute" if embedders_mode != "spin" else "mute_spin")
-
-    for name in names:
-        options.append(f"{gt_wavs_dir}/{name}.wav|{feature_dir}/{name}.npy|{f0_dir}/{name}.wav.npy|{f0nsf_dir}/{name}.wav.npy|0" if pitch_guidance else f"{gt_wavs_dir}/{name}.wav|{feature_dir}/{name}.npy|0")
-
-    mute_audio_path, mute_feature_path = os.path.join(mute_base_path, "sliced_audios", f"mute{sample_rate}.wav"), os.path.join(mute_base_path, f"{rvc_version}_extracted", "mute.npy")
-    for _ in range(2):
-        options.append(f"{mute_audio_path}|{mute_feature_path}|{os.path.join(mute_base_path, 'f0', 'mute.wav.npy')}|{os.path.join(mute_base_path, 'f0_voiced', 'mute.wav.npy')}|0" if pitch_guidance else f"{mute_audio_path}|{mute_feature_path}|0")
-
-    shuffle(options)
-    with open(os.path.join(model_path, "filelist.txt"), "w") as f:
-        f.write("\n".join(options))
-
-def setup_paths(exp_dir, version = None):
-    wav_path = os.path.join(exp_dir, "sliced_audios_16k")
-
-    if version:
-        out_path = os.path.join(exp_dir, f"{version}_extracted")
-        os.makedirs(out_path, exist_ok=True)
-        return wav_path, out_path
-    else:
-        output_root1, output_root2 = os.path.join(exp_dir, "f0"), os.path.join(exp_dir, "f0_voiced")
-        os.makedirs(output_root1, exist_ok=True); os.makedirs(output_root2, exist_ok=True)
-        return wav_path, output_root1, output_root2
-
-def read_wave(wav_path, normalize = False, is_half = False):
-    wav, sr = sf.read(wav_path, dtype=np.float32)
-    assert sr == 16000, translations["sr_not_16000"]
-
-    feats = torch.from_numpy(wav)
-    if feats.dim() == 2: feats = feats.mean(-1)
-    feats = feats.view(1, -1)
-
-    if normalize: feats = F.layer_norm(feats, feats.shape)
-    return feats.half() if is_half else feats.float()
-
-def get_device(gpu_index):
-    try:
-        index = int(gpu_index)
-        if index < torch.cuda.device_count(): return f"cuda:{index}"
-        else: logger.warning(translations["gpu_not_valid"])
-    except ValueError:
-        logger.warning(translations["gpu_not_valid"])
-        return "cpu"
-
-def get_providers():
-    ort_providers = onnxruntime.get_available_providers()
-
-    if "CUDAExecutionProvider" in ort_providers: providers = ["CUDAExecutionProvider"]
-    elif "CoreMLExecutionProvider" in ort_providers: providers = ["CoreMLExecutionProvider"]
-    else: providers = ["CPUExecutionProvider"]
-
-    return providers
-
-class FeatureInput:
-    def __init__(self, sample_rate=16000, hop_size=160, is_half=False, device=config.device):
-        self.fs = sample_rate
-        self.hop = hop_size
-        self.f0_bin = 256
-        self.f0_max = 1100.0
-        self.f0_min = 50.0
-        self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
-        self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
-        self.device = device
-        self.is_half = is_half
-    
-    def compute_f0_hybrid(self, methods_str, np_arr, hop_length, f0_onnx):
-        methods_str = re.search("hybrid\[(.+)\]", methods_str)
-        if methods_str: methods = [method.strip() for method in methods_str.group(1).split("+")]
-        f0_computation_stack, resampled_stack = [], []
-        logger.debug(translations["hybrid_methods"].format(methods=methods))
-
-        for method in methods:
-            f0 = None
-            f0_methods = {"pm": lambda: self.get_pm(np_arr), "dio": lambda: self.get_pyworld(np_arr, "dio"), "mangio-crepe-full": lambda: self.get_mangio_crepe(np_arr, int(hop_length), "full", onnx=f0_onnx), "mangio-crepe-large": lambda: self.get_mangio_crepe(np_arr, int(hop_length), "large", onnx=f0_onnx), "mangio-crepe-medium": lambda: self.get_mangio_crepe(np_arr, int(hop_length), "medium", onnx=f0_onnx), "mangio-crepe-small": lambda: self.get_mangio_crepe(np_arr, int(hop_length), "small", onnx=f0_onnx), "mangio-crepe-tiny": lambda: self.get_mangio_crepe(np_arr, int(hop_length), "tiny", onnx=f0_onnx), "crepe-full": lambda: self.get_crepe(np_arr, "full", onnx=f0_onnx), "crepe-large": lambda: self.get_crepe(np_arr, "large", onnx=f0_onnx), "crepe-medium": lambda: self.get_crepe(np_arr, "medium", onnx=f0_onnx), "crepe-small": lambda: self.get_crepe(np_arr, "small", onnx=f0_onnx), "crepe-tiny": lambda: self.get_crepe(np_arr, "tiny", onnx=f0_onnx), "fcpe": lambda: self.get_fcpe(np_arr, int(hop_length), onnx=f0_onnx), "fcpe-legacy": lambda: self.get_fcpe(np_arr, int(hop_length), legacy=True, onnx=f0_onnx), "rmvpe": lambda: self.get_rmvpe(np_arr, onnx=f0_onnx), "rmvpe-legacy": lambda: self.get_rmvpe(np_arr, legacy=True, onnx=f0_onnx), "harvest": lambda: self.get_pyworld(np_arr, "harvest"), "swipe": lambda: self.get_swipe(np_arr), "yin": lambda: self.get_yin(np_arr, int(hop_length), mode="yin"), "pyin": lambda: self.get_yin(np_arr, int(hop_length), mode="pyin")}
-            f0 = f0_methods.get(method, lambda: ValueError(translations["method_not_valid"]))()
-            f0_computation_stack.append(f0) 
-
-        for f0 in f0_computation_stack:
-            resampled_stack.append(np.interp(np.linspace(0, len(f0), (np_arr.size // self.hop)), np.arange(len(f0)), f0))
-
-        return resampled_stack[0] if len(resampled_stack) == 1 else np.nanmedian(np.vstack(resampled_stack), axis=0)
-
-    def compute_f0(self, np_arr, f0_method, hop_length, f0_onnx=False):
-        f0_methods = {"pm": lambda: self.get_pm(np_arr), "dio": lambda: self.get_pyworld(np_arr, "dio"), "mangio-crepe-full": lambda: self.get_mangio_crepe(np_arr, int(hop_length), "full", onnx=f0_onnx), "mangio-crepe-large": lambda: self.get_mangio_crepe(np_arr, int(hop_length), "large", onnx=f0_onnx), "mangio-crepe-medium": lambda: self.get_mangio_crepe(np_arr, int(hop_length), "medium", onnx=f0_onnx), "mangio-crepe-small": lambda: self.get_mangio_crepe(np_arr, int(hop_length), "small", onnx=f0_onnx), "mangio-crepe-tiny": lambda: self.get_mangio_crepe(np_arr, int(hop_length), "tiny", onnx=f0_onnx), "crepe-full": lambda: self.get_crepe(np_arr, "full", onnx=f0_onnx), "crepe-large": lambda: self.get_crepe(np_arr, "large", onnx=f0_onnx), "crepe-medium": lambda: self.get_crepe(np_arr, "medium", onnx=f0_onnx), "crepe-small": lambda: self.get_crepe(np_arr, "small", onnx=f0_onnx), "crepe-tiny": lambda: self.get_crepe(np_arr, "tiny", onnx=f0_onnx), "fcpe": lambda: self.get_fcpe(np_arr, int(hop_length), onnx=f0_onnx), "fcpe-legacy": lambda: self.get_fcpe(np_arr, int(hop_length), legacy=True, onnx=f0_onnx), "rmvpe": lambda: self.get_rmvpe(np_arr, onnx=f0_onnx), "rmvpe-legacy": lambda: self.get_rmvpe(np_arr, legacy=True, onnx=f0_onnx), "harvest": lambda: self.get_pyworld(np_arr, "harvest"), "swipe": lambda: self.get_swipe(np_arr), "yin": lambda: self.get_yin(np_arr, int(hop_length), mode="yin"), "pyin": lambda: self.get_yin(np_arr, int(hop_length), mode="pyin")}
-        return self.compute_f0_hybrid(f0_method, np_arr, int(hop_length), f0_onnx) if "hybrid" in f0_method else f0_methods.get(f0_method, lambda: ValueError(translations["method_not_valid"]))()
-
-    def get_pm(self, x):
-        import parselmouth
-
-        f0 = (parselmouth.Sound(x, self.fs).to_pitch_ac(time_step=(160 / 16000 * 1000) / 1000, voicing_threshold=0.6, pitch_floor=50, pitch_ceiling=1100).selected_array["frequency"])
-        pad_size = ((x.size // self.hop) - len(f0) + 1) // 2
-
-        if pad_size > 0 or (x.size // self.hop) - len(f0) - pad_size > 0: f0 = np.pad(f0, [[pad_size, (x.size // self.hop) - len(f0) - pad_size]], mode="constant")
-        return f0
-    
-    def get_mangio_crepe(self, x, hop_length, model="full", onnx=False):
-        from main.library.predictors.CREPE import predict
-
-        audio = torch.from_numpy(x.astype(np.float32)).to(self.device)
-        audio /= torch.quantile(torch.abs(audio), 0.999)
-        audio = audio.unsqueeze(0)
-        source = predict(audio, self.fs, hop_length, self.f0_min, self.f0_max, model=model, batch_size=hop_length * 2, device=self.device, pad=True, providers=get_providers(), onnx=onnx).squeeze(0).cpu().float().numpy()
-        source[source < 0.001] = np.nan
-
-        return np.nan_to_num(np.interp(np.arange(0, len(source) * (x.size // self.hop), len(source)) / (x.size // self.hop), np.arange(0, len(source)), source))
-    
-    def get_crepe(self, x, model="full", onnx=False):
-        from main.library.predictors.CREPE import predict, mean, median
-
-        f0, pd = predict(torch.tensor(np.copy(x))[None].float(), self.fs, 160, self.f0_min, self.f0_max, model, batch_size=512, device=self.device, return_periodicity=True, providers=get_providers(), onnx=onnx)
-        f0, pd = mean(f0, 3), median(pd, 3)
-        f0[pd < 0.1] = 0
-
-        return f0[0].cpu().numpy()
-    
-    def get_fcpe(self, x, hop_length, legacy=False, onnx=False):
-        from main.library.predictors.FCPE import FCPE
-
-        model_fcpe = FCPE(os.path.join("assets", "models", "predictors", ("fcpe_legacy" if legacy else"fcpe") + (".onnx" if onnx else ".pt")), hop_length=int(hop_length), f0_min=int(self.f0_min), f0_max=int(self.f0_max), dtype=torch.float32, device=self.device, sample_rate=self.fs, threshold=0.03 if legacy else 0.006, providers=get_providers(), onnx=onnx, legacy=legacy)
-        f0 = model_fcpe.compute_f0(x, p_len=(x.size // self.hop))
-
-        del model_fcpe
-        return f0
-    
-    def get_rmvpe(self, x, legacy=False, onnx=False):
-        from main.library.predictors.RMVPE import RMVPE
-
-        rmvpe_model = RMVPE(os.path.join("assets", "models", "predictors", "rmvpe" + (".onnx" if onnx else ".pt")), is_half=self.is_half, device=self.device, onnx=onnx, providers=get_providers())
-        f0 = rmvpe_model.infer_from_audio_with_pitch(x, thred=0.03, f0_min=self.f0_min, f0_max=self.f0_max) if legacy else rmvpe_model.infer_from_audio(x, thred=0.03)
-
-        del rmvpe_model
-        return f0
-    
-    def get_pyworld(self, x, model="harvest"):
-        from main.library.predictors.WORLD_WRAPPER import PYWORLD
-
-        pw = PYWORLD()
-        x = x.astype(np.double)
-
-        if model == "harvest": f0, t = pw.harvest(x, fs=self.fs, f0_ceil=self.f0_max, f0_floor=self.f0_min, frame_period=1000 * self.hop / self.fs)
-        elif model == "dio": f0, t = pw.dio(x, fs=self.fs, f0_ceil=self.f0_max, f0_floor=self.f0_min, frame_period=1000 * self.hop / self.fs)
-        else: raise ValueError(translations["method_not_valid"])
-
-        return pw.stonemask(x, self.fs, t, f0)
-    
-    def get_swipe(self, x):
-        from main.library.predictors.SWIPE import swipe
-
-        f0, _ = swipe(x.astype(np.float32), self.fs, f0_floor=self.f0_min, f0_ceil=self.f0_max, frame_period=1000 * self.hop / self.fs)
-        return f0
-    
-    def get_yin(self, x, hop_length, mode="yin"):
-        import librosa
-
-        source = np.array(librosa.yin(x.astype(np.float32), sr=self.fs, fmin=self.f0_min, fmax=self.f0_max, hop_length=hop_length) if mode == "yin" else librosa.pyin(x.astype(np.float32), fmin=self.f0_min, fmax=self.f0_max, sr=self.fs, hop_length=hop_length)[0])
-        source[source < 0.001] = np.nan
-        return np.nan_to_num(np.interp(np.arange(0, len(source) * (x.size // self.hop), len(source)) / (x.size // self.hop), np.arange(0, len(source)), source))
-    
-    def coarse_f0(self, f0):
-        return np.rint(np.clip(((1127 * np.log(1 + f0 / 700)) - self.f0_mel_min) * (self.f0_bin - 2) / (self.f0_mel_max - self.f0_mel_min) + 1, 1, self.f0_bin - 1)).astype(int)
-
-    def process_file(self, file_info, f0_method, hop_length, f0_onnx):
-        inp_path, opt_path1, opt_path2, np_arr = file_info
-        if os.path.exists(opt_path1 + ".npy") and os.path.exists(opt_path2 + ".npy"): return
-
-        try:
-            feature_pit = self.compute_f0(np_arr, f0_method, hop_length, f0_onnx)
-            if isinstance(feature_pit, tuple): feature_pit = feature_pit[0]
-            np.save(opt_path2, feature_pit, allow_pickle=False)
-            np.save(opt_path1, self.coarse_f0(feature_pit), allow_pickle=False)
-        except Exception as e:
-            raise RuntimeError(f"{translations['extract_file_error']} {inp_path}: {e}")
-
-    def process_files(self, files, f0_method, hop_length, f0_onnx, pbar):
-        for file_info in files:
-            self.process_file(file_info, f0_method, hop_length, f0_onnx)
-            pbar.update()
-
-def run_pitch_extraction(exp_dir, f0_method, hop_length, num_processes, gpus, f0_onnx, is_half):
-    input_root, *output_roots = setup_paths(exp_dir)
-    output_root1, output_root2 = output_roots if len(output_roots) == 2 else (output_roots[0], None)
-
-    paths = [(os.path.join(input_root, name), os.path.join(output_root1, name) if output_root1 else None, os.path.join(output_root2, name) if output_root2 else None, load_audio(logger, os.path.join(input_root, name), 16000)) for name in sorted(os.listdir(input_root)) if "spec" not in name]
-    logger.info(translations["extract_f0_method"].format(num_processes=num_processes, f0_method=f0_method))
-
-    start_time = time.time()
-    gpus = gpus.split("-")
-    process_partials = []
-
-    pbar = tqdm.tqdm(total=len(paths), ncols=100, unit="p")
-    for idx, gpu in enumerate(gpus):
-        feature_input = FeatureInput(device=get_device(gpu) if gpu != "" else "cpu", is_half=is_half)
-        process_partials.append((feature_input, paths[idx::len(gpus)]))
-
-    with ThreadPoolExecutor(max_workers=num_processes) as executor:
-        for future in as_completed([executor.submit(FeatureInput.process_files, feature_input, part_paths, f0_method, hop_length, f0_onnx, pbar) for feature_input, part_paths in process_partials]):
-            pbar.update(1)
-            logger.debug(pbar.format_meter(pbar.n, pbar.total, pbar.format_dict["elapsed"]))
-            future.result()
-
-    pbar.close()
-    logger.info(translations["extract_f0_success"].format(elapsed_time=f"{(time.time() - start_time):.2f}"))
-
-def extract_features(model, feats, version):
-    return torch.as_tensor(model.run([model.get_outputs()[0].name, model.get_outputs()[1].name], {"feats": feats.detach().cpu().numpy()})[0 if version == "v1" else 1], dtype=torch.float32, device=feats.device)
-
-def process_file_embedding(file, wav_path, out_path, model, device, version, saved_cfg, embed_suffix, is_half):
-    out_file_path = os.path.join(out_path, file.replace("wav", "npy"))
-    if os.path.exists(out_file_path): return
-    feats = read_wave(os.path.join(wav_path, file), normalize=saved_cfg.task.normalize if saved_cfg else False, is_half=is_half).to(device)
-
-    with torch.no_grad():
-        if embed_suffix == ".pt":
-            model = model.to(device).to(torch.float16 if is_half else torch.float32).eval()
-            logits = model.extract_features(**{"source": feats, "padding_mask": torch.BoolTensor(feats.shape).fill_(False).to(device), "output_layer": 9 if version == "v1" else 12})
-            feats = model.final_proj(logits[0]) if version == "v1" else logits[0]
-        elif embed_suffix == ".onnx": feats = extract_features(model, feats, version).to(device)
-        elif embed_suffix == ".safetensors":
-            model = model.to(device).to(torch.float16 if is_half else torch.float32).eval()
-            logits = model(feats)["last_hidden_state"]
-            feats = (model.final_proj(logits[0]).unsqueeze(0) if version == "v1" else logits)
-        else: raise ValueError(translations["option_not_valid"])
-
-    feats = feats.squeeze(0).float().cpu().numpy()
-    if not np.isnan(feats).any(): np.save(out_file_path, feats, allow_pickle=False)
-    else: logger.warning(f"{file} {translations['NaN']}")
-
-def run_embedding_extraction(exp_dir, version, gpus, embedder_model, embedders_mode, is_half):
-    wav_path, out_path = setup_paths(exp_dir, version)
-    logger.info(translations["start_extract_hubert"])
-    start_time = time.time()
-    models, saved_cfg, embed_suffix = load_embedders_model(embedder_model, embedders_mode, providers=get_providers())
-    devices = [get_device(gpu) for gpu in (gpus.split("-") if gpus != "-" else ["cpu"])]
-    paths = sorted([file for file in os.listdir(wav_path) if file.endswith(".wav")])
-
-    if not paths:
-        logger.warning(translations["not_found_audio_file"])
-        sys.exit(1)
-
-    pbar = tqdm.tqdm(total=len(paths) * len(devices), ncols=100, unit="p")
-    for task in [(file, wav_path, out_path, models, device, version, saved_cfg, embed_suffix, is_half) for file in paths for device in devices]:
-        try:
-            process_file_embedding(*task)
-        except Exception as e:
-            raise RuntimeError(f"{translations['process_error']} {task[0]}: {e}")
-        
-        pbar.update(1)
-        logger.debug(pbar.format_meter(pbar.n, pbar.total, pbar.format_dict["elapsed"]))
-
-    pbar.close()
-    logger.info(translations["extract_hubert_success"].format(elapsed_time=f"{(time.time() - start_time):.2f}"))
-
-def main():
-    args = parse_arguments()
-    exp_dir = os.path.join("assets", "logs", args.model_name)
-    f0_method, hop_length, num_processes, gpus, version, pitch_guidance, sample_rate, embedder_model, f0_onnx, embedders_mode = args.f0_method, args.hop_length, args.cpu_cores, args.gpu, args.rvc_version, args.pitch_guidance, args.sample_rate, args.embedder_model, args.f0_onnx, args.embedders_mode
-
-    check_predictors(f0_method, f0_onnx); check_embedders(embedder_model, embedders_mode)
-    if logger.hasHandlers(): logger.handlers.clear()
-    else:
-        console_handler = logging.StreamHandler()
-        console_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
-        console_handler.setFormatter(console_formatter)
-        console_handler.setLevel(logging.INFO)
-        file_handler = logging.handlers.RotatingFileHandler(os.path.join(exp_dir, "extract.log"), maxBytes=5*1024*1024, backupCount=3, encoding='utf-8')
-        file_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
-        file_handler.setFormatter(file_formatter)
-        file_handler.setLevel(logging.DEBUG)
-        logger.addHandler(console_handler)
-        logger.addHandler(file_handler)
-        logger.setLevel(logging.DEBUG)
-
-    log_data = {translations['modelname']: args.model_name, translations['export_process']: exp_dir, translations['f0_method']: f0_method, translations['pretrain_sr']: sample_rate, translations['cpu_core']: num_processes, "Gpu": gpus, "Hop length": hop_length, translations['training_version']: version, translations['extract_f0']: pitch_guidance, translations['hubert_model']: embedder_model, translations["f0_onnx_mode"]: f0_onnx, translations["embed_mode"]: embedders_mode}
-    for key, value in log_data.items():
-        logger.debug(f"{key}: {value}")
-
-    pid_path = os.path.join(exp_dir, "extract_pid.txt")
-    with open(pid_path, "w") as pid_file:
-        pid_file.write(str(os.getpid()))
-
-    try:
-        run_pitch_extraction(exp_dir, f0_method, hop_length, num_processes, gpus, f0_onnx, config.is_half)
-        run_embedding_extraction(exp_dir, version, gpus, embedder_model, embedders_mode, config.is_half)
-        generate_config(version, sample_rate, exp_dir)
-        generate_filelist(pitch_guidance, exp_dir, version, sample_rate, embedders_mode)
-    except Exception as e:
-        logger.error(f"{translations['extract_error']}: {e}")
-        import traceback
-        logger.debug(traceback.format_exc())
-
-    if os.path.exists(pid_path): os.remove(pid_path)
-    logger.info(f"{translations['extract_success']} {args.model_name}.")
-
-if __name__ == "__main__": main()

main/inference/preprocess.py

@@ -1,270 +0,0 @@
-import os
-import sys
-import time
-import logging
-import librosa
-import argparse
-import logging.handlers
-
-import numpy as np
-import torch.multiprocessing as mp
-
-from tqdm import tqdm
-from scipy import signal
-from scipy.io import wavfile
-from distutils.util import strtobool
-from concurrent.futures import ProcessPoolExecutor, as_completed
-
-sys.path.append(os.getcwd())
-
-from main.library.utils import load_audio
-from main.configs.config import Config
-
-logger = logging.getLogger(__name__)
-for l in ["numba.core.byteflow", "numba.core.ssa", "numba.core.interpreter"]:
-    logging.getLogger(l).setLevel(logging.ERROR)
-
-OVERLAP, MAX_AMPLITUDE, ALPHA, HIGH_PASS_CUTOFF, SAMPLE_RATE_16K = 0.3, 0.9, 0.75, 48, 16000
-
-config = Config()
-translations = config.translations
-
-def parse_arguments():
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--model_name", type=str, required=True)
-    parser.add_argument("--dataset_path", type=str, default="./dataset")
-    parser.add_argument("--sample_rate", type=int, required=True)
-    parser.add_argument("--cpu_cores", type=int, default=2)
-    parser.add_argument("--cut_preprocess", type=lambda x: bool(strtobool(x)), default=True)
-    parser.add_argument("--process_effects", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--clean_dataset", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--clean_strength", type=float, default=0.7)
-
-    return parser.parse_args()
-
-class Slicer:
-    def __init__(self, sr, threshold = -40.0, min_length = 5000, min_interval = 300, hop_size = 20, max_sil_kept = 5000):
-        if not min_length >= min_interval >= hop_size: raise ValueError(translations["min_length>=min_interval>=hop_size"])
-        if not max_sil_kept >= hop_size: raise ValueError(translations["max_sil_kept>=hop_size"])
-        min_interval = sr * min_interval / 1000
-        self.threshold = 10 ** (threshold / 20.0)
-        self.hop_size = round(sr * hop_size / 1000)
-        self.win_size = min(round(min_interval), 4 * self.hop_size)
-        self.min_length = round(sr * min_length / 1000 / self.hop_size)
-        self.min_interval = round(min_interval / self.hop_size)
-        self.max_sil_kept = round(sr * max_sil_kept / 1000 / self.hop_size)
-
-    def _apply_slice(self, waveform, begin, end):
-        start_idx = begin * self.hop_size
-
-        if len(waveform.shape) > 1: return waveform[:, start_idx:min(waveform.shape[1], end * self.hop_size)]
-        else: return waveform[start_idx:min(waveform.shape[0], end * self.hop_size)]
-
-    def slice(self, waveform):
-        samples = waveform.mean(axis=0) if len(waveform.shape) > 1 else waveform
-        if samples.shape[0] <= self.min_length: return [waveform]
-        rms_list = get_rms(y=samples, frame_length=self.win_size, hop_length=self.hop_size).squeeze(0)
-        sil_tags = []
-        silence_start, clip_start = None, 0
-
-        for i, rms in enumerate(rms_list):
-            if rms < self.threshold:
-                if silence_start is None: silence_start = i
-                continue
-
-            if silence_start is None: continue
-
-            is_leading_silence = silence_start == 0 and i > self.max_sil_kept
-            need_slice_middle = (i - silence_start >= self.min_interval and i - clip_start >= self.min_length)
-
-            if not is_leading_silence and not need_slice_middle:
-                silence_start = None
-                continue
-
-            if i - silence_start <= self.max_sil_kept:
-                pos = rms_list[silence_start : i + 1].argmin() + silence_start
-                sil_tags.append((0, pos) if silence_start == 0 else (pos, pos))   
-                clip_start = pos
-            elif i - silence_start <= self.max_sil_kept * 2:
-                pos = rms_list[i - self.max_sil_kept : silence_start + self.max_sil_kept + 1].argmin()
-                pos += i - self.max_sil_kept
-                pos_r = (rms_list[i - self.max_sil_kept : i + 1].argmin() + i - self.max_sil_kept)
-
-                if silence_start == 0:
-                    sil_tags.append((0, pos_r))
-                    clip_start = pos_r
-                else:
-                    sil_tags.append((min((rms_list[silence_start : silence_start + self.max_sil_kept + 1].argmin() + silence_start), pos), max(pos_r, pos)))
-                    clip_start = max(pos_r, pos)
-            else:
-                pos_r = (rms_list[i - self.max_sil_kept : i + 1].argmin() + i - self.max_sil_kept)
-                sil_tags.append((0, pos_r) if silence_start == 0 else ((rms_list[silence_start : silence_start + self.max_sil_kept + 1].argmin() + silence_start), pos_r))
-                clip_start = pos_r
-
-            silence_start = None
-        total_frames = rms_list.shape[0]
-        if (silence_start is not None and total_frames - silence_start >= self.min_interval): sil_tags.append((rms_list[silence_start : min(total_frames, silence_start + self.max_sil_kept) + 1].argmin() + silence_start, total_frames + 1))
-
-        if not sil_tags: return [waveform]
-        else:
-            chunks = []
-            if sil_tags[0][0] > 0: chunks.append(self._apply_slice(waveform, 0, sil_tags[0][0]))
-
-            for i in range(len(sil_tags) - 1):
-                chunks.append(self._apply_slice(waveform, sil_tags[i][1], sil_tags[i + 1][0]))
-
-            if sil_tags[-1][1] < total_frames: chunks.append(self._apply_slice(waveform, sil_tags[-1][1], total_frames))
-            return chunks
-
-def get_rms(y, frame_length=2048, hop_length=512, pad_mode="constant"):
-    y = np.pad(y, (int(frame_length // 2), int(frame_length // 2)), mode=pad_mode)
-    axis = -1
-    x_shape_trimmed = list(y.shape)
-    x_shape_trimmed[axis] -= frame_length - 1
-    xw = np.moveaxis(np.lib.stride_tricks.as_strided(y, shape=tuple(x_shape_trimmed) + tuple([frame_length]), strides=y.strides + tuple([y.strides[axis]])), -1, axis - 1 if axis < 0 else axis + 1)
-    slices = [slice(None)] * xw.ndim
-    slices[axis] = slice(0, None, hop_length)
-    return np.sqrt(np.mean(np.abs(xw[tuple(slices)]) ** 2, axis=-2, keepdims=True))
-
-class PreProcess:
-    def __init__(self, sr, exp_dir, per):
-        self.slicer = Slicer(sr=sr, threshold=-42, min_length=1500, min_interval=400, hop_size=15, max_sil_kept=500)
-        self.sr = sr
-        self.b_high, self.a_high = signal.butter(N=5, Wn=HIGH_PASS_CUTOFF, btype="high", fs=self.sr)
-        self.per = per
-        self.exp_dir = exp_dir
-        self.device = "cpu"
-        self.gt_wavs_dir = os.path.join(exp_dir, "sliced_audios")
-        self.wavs16k_dir = os.path.join(exp_dir, "sliced_audios_16k")
-        os.makedirs(self.gt_wavs_dir, exist_ok=True)
-        os.makedirs(self.wavs16k_dir, exist_ok=True)
-
-    def _normalize_audio(self, audio):
-        tmp_max = np.abs(audio).max()
-        if tmp_max > 2.5: return None
-        return (audio / tmp_max * (MAX_AMPLITUDE * ALPHA)) + (1 - ALPHA) * audio
-
-    def process_audio_segment(self, normalized_audio, sid, idx0, idx1):
-        if normalized_audio is None:
-            logger.debug(f"{sid}-{idx0}-{idx1}-filtered")
-            return
-        
-        wavfile.write(os.path.join(self.gt_wavs_dir, f"{sid}_{idx0}_{idx1}.wav"), self.sr, normalized_audio.astype(np.float32))
-        wavfile.write(os.path.join(self.wavs16k_dir, f"{sid}_{idx0}_{idx1}.wav"), SAMPLE_RATE_16K, librosa.resample(normalized_audio, orig_sr=self.sr, target_sr=SAMPLE_RATE_16K, res_type="soxr_vhq").astype(np.float32))
-
-    def process_audio(self, path, idx0, sid, cut_preprocess, process_effects, clean_dataset, clean_strength):
-        try:
-            audio = load_audio(logger, path, self.sr)
-
-            if process_effects: 
-                audio = signal.lfilter(self.b_high, self.a_high, audio)
-                audio = self._normalize_audio(audio)
-
-            if clean_dataset: 
-                from main.tools.noisereduce import reduce_noise
-                audio = reduce_noise(y=audio, sr=self.sr, prop_decrease=clean_strength, device=config.device)
-
-            idx1 = 0
-            if cut_preprocess:
-                for audio_segment in self.slicer.slice(audio):
-                    i = 0
-
-                    while 1:
-                        start = int(self.sr * (self.per - OVERLAP) * i)
-                        i += 1
-
-                        if len(audio_segment[start:]) > (self.per + OVERLAP) * self.sr:
-                            self.process_audio_segment(audio_segment[start : start + int(self.per * self.sr)], sid, idx0, idx1)
-                            idx1 += 1
-                        else:
-                            self.process_audio_segment(audio_segment[start:], sid, idx0, idx1)
-                            idx1 += 1
-                            break
-            else: self.process_audio_segment(audio, sid, idx0, idx1)
-        except Exception as e:
-            raise RuntimeError(f"{translations['process_audio_error']}: {e}")
-
-def process_file(args):
-    pp, file, cut_preprocess, process_effects, clean_dataset, clean_strength = (args)
-    file_path, idx0, sid = file
-    pp.process_audio(file_path, idx0, sid, cut_preprocess, process_effects, clean_dataset, clean_strength)
-
-def preprocess_training_set(input_root, sr, num_processes, exp_dir, per, cut_preprocess, process_effects, clean_dataset, clean_strength):
-    start_time = time.time()
-    pp = PreProcess(sr, exp_dir, per)
-    logger.info(translations["start_preprocess"].format(num_processes=num_processes))
-    files = []
-    idx = 0
-
-    for root, _, filenames in os.walk(input_root):
-        try:
-            sid = 0 if root == input_root else int(os.path.basename(root))
-
-            for f in filenames:
-                if f.lower().endswith(("wav", "mp3", "flac", "ogg", "opus", "m4a", "mp4", "aac", "alac", "wma", "aiff", "webm", "ac3")):
-                    files.append((os.path.join(root, f), idx, sid))
-                    idx += 1
-        except ValueError:
-            raise ValueError(f"{translations['not_integer']} '{os.path.basename(root)}'.")
-
-    with tqdm(total=len(files), ncols=100, unit="f") as pbar:
-        with ProcessPoolExecutor(max_workers=num_processes) as executor:
-            futures = [executor.submit(process_file, (pp, file, cut_preprocess, process_effects, clean_dataset, clean_strength)) for file in files]
-            for future in as_completed(futures):
-                try:
-                    future.result() 
-                except Exception as e:
-                    raise RuntimeError(f"{translations['process_error']}: {e}")
-                pbar.update(1)
-                logger.debug(pbar.format_meter(pbar.n, pbar.total, pbar.format_dict["elapsed"]))
-
-    elapsed_time = time.time() - start_time
-    logger.info(translations["preprocess_success"].format(elapsed_time=f"{elapsed_time:.2f}"))
-
-def main():
-    args = parse_arguments()
-    experiment_directory = os.path.join("assets", "logs", args.model_name)
-
-    num_processes = args.cpu_cores
-    num_processes = 2 if num_processes is None else int(num_processes)
-    dataset, sample_rate, cut_preprocess, preprocess_effects, clean_dataset, clean_strength = args.dataset_path, args.sample_rate, args.cut_preprocess, args.process_effects, args.clean_dataset, args.clean_strength
-
-    os.makedirs(experiment_directory, exist_ok=True)
-    
-    if logger.hasHandlers(): logger.handlers.clear()
-    else:
-        console_handler = logging.StreamHandler()
-        console_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
-        console_handler.setFormatter(console_formatter)
-        console_handler.setLevel(logging.INFO)
-        file_handler = logging.handlers.RotatingFileHandler(os.path.join(experiment_directory, "preprocess.log"), maxBytes=5*1024*1024, backupCount=3, encoding='utf-8')
-        file_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
-        file_handler.setFormatter(file_formatter)
-        file_handler.setLevel(logging.DEBUG)
-        logger.addHandler(console_handler)
-        logger.addHandler(file_handler)
-        logger.setLevel(logging.DEBUG)
-
-    log_data = {translations['modelname']: args.model_name, translations['export_process']: experiment_directory, translations['dataset_folder']: dataset, translations['pretrain_sr']: sample_rate, translations['cpu_core']: num_processes, translations['split_audio']: cut_preprocess, translations['preprocess_effect']: preprocess_effects, translations['clear_audio']: clean_dataset}
-    if clean_dataset: log_data[translations['clean_strength']] = clean_strength
-
-    for key, value in log_data.items():
-        logger.debug(f"{key}: {value}")
-
-    pid_path = os.path.join(experiment_directory, "preprocess_pid.txt")
-    with open(pid_path, "w") as pid_file:
-        pid_file.write(str(os.getpid()))
-    
-    try:
-        preprocess_training_set(dataset, sample_rate, num_processes, experiment_directory, config.per_preprocess, cut_preprocess, preprocess_effects, clean_dataset, clean_strength)
-    except Exception as e:
-        logger.error(f"{translations['process_audio_error']} {e}")
-        import traceback
-        logger.debug(traceback.format_exc())
-        
-    if os.path.exists(pid_path): os.remove(pid_path)
-    logger.info(f"{translations['preprocess_model_success']} {args.model_name}")
-
-if __name__ == "__main__": 
-    mp.set_start_method("spawn", force=True)
-    main()

main/inference/separator_music.py

@@ -1,310 +0,0 @@
-import os
-import sys
-import time
-import logging
-import argparse
-import logging.handlers
-
-import numpy as np
-
-from distutils.util import strtobool
-
-sys.path.append(os.getcwd())
-
-from main.configs.config import Config
-from main.library.algorithm.separator import Separator
-from main.library.utils import pydub_convert, pydub_load
-
-config = Config()
-translations = config.translations 
-logger = logging.getLogger(__name__)
-
-if logger.hasHandlers(): logger.handlers.clear()
-else:
-    console_handler = logging.StreamHandler()
-    console_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
-    console_handler.setFormatter(console_formatter)
-    console_handler.setLevel(logging.INFO)
-    file_handler = logging.handlers.RotatingFileHandler(os.path.join("assets", "logs", "separator.log"), maxBytes=5*1024*1024, backupCount=3, encoding='utf-8')
-    file_formatter = logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S")
-    file_handler.setFormatter(file_formatter)
-    file_handler.setLevel(logging.DEBUG)
-    logger.addHandler(console_handler)
-    logger.addHandler(file_handler)
-    logger.setLevel(logging.DEBUG)
-
-demucs_models = {"HT-Tuned": "htdemucs_ft.yaml", "HT-Normal": "htdemucs.yaml", "HD_MMI": "hdemucs_mmi.yaml", "HT_6S": "htdemucs_6s.yaml"}
-mdx_models = {"Main_340": "UVR-MDX-NET_Main_340.onnx", "Main_390": "UVR-MDX-NET_Main_390.onnx", "Main_406": "UVR-MDX-NET_Main_406.onnx", "Main_427": "UVR-MDX-NET_Main_427.onnx","Main_438": "UVR-MDX-NET_Main_438.onnx", "Inst_full_292": "UVR-MDX-NET-Inst_full_292.onnx", "Inst_HQ_1": "UVR-MDX-NET-Inst_HQ_1.onnx", "Inst_HQ_2": "UVR-MDX-NET-Inst_HQ_2.onnx", "Inst_HQ_3": "UVR-MDX-NET-Inst_HQ_3.onnx", "Inst_HQ_4": "UVR-MDX-NET-Inst_HQ_4.onnx", "Inst_HQ_5": "UVR-MDX-NET-Inst_HQ_5.onnx", "Kim_Vocal_1": "Kim_Vocal_1.onnx", "Kim_Vocal_2": "Kim_Vocal_2.onnx", "Kim_Inst": "Kim_Inst.onnx", "Inst_187_beta": "UVR-MDX-NET_Inst_187_beta.onnx", "Inst_82_beta": "UVR-MDX-NET_Inst_82_beta.onnx", "Inst_90_beta": "UVR-MDX-NET_Inst_90_beta.onnx", "Voc_FT": "UVR-MDX-NET-Voc_FT.onnx", "Crowd_HQ": "UVR-MDX-NET_Crowd_HQ_1.onnx", "MDXNET_9482": "UVR_MDXNET_9482.onnx", "Inst_1": "UVR-MDX-NET-Inst_1.onnx", "Inst_2": "UVR-MDX-NET-Inst_2.onnx", "Inst_3": "UVR-MDX-NET-Inst_3.onnx", "MDXNET_1_9703": "UVR_MDXNET_1_9703.onnx", "MDXNET_2_9682": "UVR_MDXNET_2_9682.onnx", "MDXNET_3_9662": "UVR_MDXNET_3_9662.onnx", "Inst_Main": "UVR-MDX-NET-Inst_Main.onnx", "MDXNET_Main": "UVR_MDXNET_Main.onnx"}
-kara_models = {"Version-1": "UVR_MDXNET_KARA.onnx", "Version-2": "UVR_MDXNET_KARA_2.onnx"}
-
-def parse_arguments():
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--input_path", type=str, required=True)
-    parser.add_argument("--output_path", type=str, default="./audios")
-    parser.add_argument("--format", type=str, default="wav")
-    parser.add_argument("--shifts", type=int, default=2)
-    parser.add_argument("--segments_size", type=int, default=256)
-    parser.add_argument("--overlap", type=float, default=0.25)
-    parser.add_argument("--mdx_hop_length", type=int, default=1024)
-    parser.add_argument("--mdx_batch_size", type=int, default=1)
-    parser.add_argument("--clean_audio", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--clean_strength", type=float, default=0.7)
-    parser.add_argument("--model_name", type=str, default="HT-Normal")
-    parser.add_argument("--kara_model", type=str, default="Version-1")
-    parser.add_argument("--backing", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--mdx_denoise", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--reverb", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--backing_reverb", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--sample_rate", type=int, default=44100)
-
-    return parser.parse_args()
-
-def main():
-    start_time = time.time()
-    pid_path = os.path.join("assets", "separate_pid.txt")
-
-    with open(pid_path, "w") as pid_file:
-        pid_file.write(str(os.getpid()))
-
-    try:
-        args = parse_arguments()
-        input_path, output_path, export_format, shifts, segments_size, overlap, hop_length, batch_size, clean_audio, clean_strength, model_name, kara_model, backing, mdx_denoise, reverb, backing_reverb, sample_rate = args.input_path, args.output_path, args.format, args.shifts, args.segments_size, args.overlap, args.mdx_hop_length, args.mdx_batch_size, args.clean_audio, args.clean_strength, args.model_name, args.kara_model, args.backing, args.mdx_denoise, args.reverb, args.backing_reverb, args.sample_rate
-
-        if backing_reverb and not reverb: 
-            logger.warning(translations["turn_on_dereverb"])
-            sys.exit(1)
-
-        if backing_reverb and not backing: 
-            logger.warning(translations["turn_on_separator_backing"])
-            sys.exit(1)
-
-        input_path = input_path.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
-        output_path = os.path.dirname(output_path) or output_path
-
-        log_data = {translations['audio_path']: input_path, translations['output_path']: output_path, translations['export_format']: export_format, translations['shift']: shifts, translations['segments_size']: segments_size, translations['overlap']: overlap, translations['modelname']: model_name, translations['denoise_mdx']: mdx_denoise, "Hop length": hop_length, translations['batch_size']: batch_size, translations['sr']: sample_rate}
-
-        if clean_audio:
-            log_data[translations['clear_audio']] = clean_audio
-            log_data[translations['clean_strength']] = clean_strength
-
-        if backing:
-            log_data[translations['backing_model_ver']] = kara_model
-            log_data[translations['separator_backing']] = backing
-
-        if reverb:
-            log_data[translations['dereveb_audio']] = reverb
-            log_data[translations['dereveb_backing']] = backing_reverb
-
-        for key, value in log_data.items():
-            logger.debug(f"{key}: {value}")
-
-        if os.path.isdir(input_path):
-            for f in input_path:
-                separation(f, output_path, export_format, shifts, overlap, segments_size, model_name, sample_rate, mdx_denoise, hop_length, batch_size, backing, reverb, kara_model, backing_reverb, clean_audio, clean_strength)
-        else: separation(input_path, output_path, export_format, shifts, overlap, segments_size, model_name, sample_rate, mdx_denoise, hop_length, batch_size, backing, reverb, kara_model, backing_reverb, clean_audio, clean_strength)
-
-    except Exception as e:
-        logger.error(f"{translations['separator_error']}: {e}")
-        import traceback
-        logger.debug(traceback.format_exc())
-    
-    if os.path.exists(pid_path): os.remove(pid_path)
-    elapsed_time = time.time() - start_time
-    logger.info(translations["separator_success"].format(elapsed_time=f"{elapsed_time:.2f}"))
-
-def separation(input_path, output_path, export_format, shifts, overlap, segments_size, model_name, sample_rate, mdx_denoise, hop_length, batch_size, backing, reverb, kara_model, backing_reverb, clean_audio, clean_strength):
-    filename, _ = os.path.splitext(os.path.basename(input_path))
-    output_path = os.path.join(output_path, filename)
-    os.makedirs(output_path, exist_ok=True)
-
-    if model_name in ["HT-Tuned", "HT-Normal", "HD_MMI", "HT_6S"]: vocals, _ = separator_music_demucs(input_path, output_path, export_format, shifts, overlap, segments_size, model_name, sample_rate)
-    else: vocals, _ = separator_music_mdx(input_path, output_path, export_format, segments_size, overlap, mdx_denoise, model_name, hop_length, batch_size, sample_rate)
-
-    if backing: main_vocals, backing_vocals = separator_backing(vocals, output_path, export_format, segments_size, overlap, mdx_denoise, kara_model, hop_length, batch_size, sample_rate)
-    if reverb: vocals_no_reverb, main_vocals_no_reverb, backing_vocals_no_reverb = separator_reverb(output_path, export_format, segments_size, overlap, mdx_denoise, reverb, backing_reverb, hop_length, batch_size, sample_rate)
-
-    original_output = os.path.join(output_path, f"Original_Vocals_No_Reverb.{export_format}") if reverb else os.path.join(output_path, f"Original_Vocals.{export_format}")
-    main_output = os.path.join(output_path, f"Main_Vocals_No_Reverb.{export_format}") if reverb and backing_reverb else os.path.join(output_path, f"Main_Vocals.{export_format}")
-    backing_output = os.path.join(output_path, f"Backing_Vocals_No_Reverb.{export_format}") if reverb and backing_reverb else os.path.join(output_path, f"Backing_Vocals.{export_format}")
-    
-    if clean_audio:
-        import soundfile as sf
-        
-        logger.info(f"{translations['clear_audio']}...")
-        vocal_data, vocal_sr = sf.read(vocals_no_reverb if reverb else vocals, dtype=np.float32)
-
-        from main.tools.noisereduce import reduce_noise
-        sf.write(original_output, reduce_noise(y=vocal_data, sr=vocal_sr, prop_decrease=clean_strength), vocal_sr, format=export_format, device=config.device)
-
-        if backing:
-            main_data, main_sr = sf.read(main_vocals_no_reverb if reverb and backing else main_vocals, dtype=np.float32)
-            backing_data, backing_sr = sf.read(backing_vocals_no_reverb if reverb and backing_reverb else backing_vocals, dtype=np.float32)
-
-            sf.write(main_output, reduce_noise(y=main_data, sr=main_sr, prop_decrease=clean_strength), main_sr, format=export_format, device=config.device)
-            sf.write(backing_output, reduce_noise(y=backing_data, sr=backing_sr, prop_decrease=clean_strength), backing_sr, format=export_format, device=config.device)  
-
-        logger.info(translations["clean_audio_success"])
-        
-def separator_music_demucs(input, output, format, shifts, overlap, segments_size, demucs_model, sample_rate):
-    if not os.path.exists(input): 
-        logger.warning(translations["input_not_valid"])
-        sys.exit(1)
-    
-    if not os.path.exists(output): 
-        logger.warning(translations["output_not_valid"])
-        sys.exit(1)
-    
-    for i in [f"Original_Vocals.{format}", f"Instruments.{format}"]:
-        if os.path.exists(os.path.join(output, i)): os.remove(os.path.join(output, i))
-
-    logger.info(f"{translations['separator_process_2']}...")
-    demucs_output = separator_main(audio_file=input, model_filename=demucs_models.get(demucs_model), output_format=format, output_dir=output, demucs_segment_size=(segments_size / 2), demucs_shifts=shifts, demucs_overlap=overlap, sample_rate=sample_rate)
-    
-    for f in demucs_output:
-        path = os.path.join(output, f)
-        if not os.path.exists(path): logger.error(translations["not_found"].format(name=path))
-
-        if '_(Drums)_' in f: drums = path
-        elif '_(Bass)_' in f: bass = path
-        elif '_(Other)_' in f: other = path
-        elif '_(Vocals)_' in f: os.rename(path, os.path.join(output, f"Original_Vocals.{format}"))
-
-    pydub_convert(pydub_load(drums)).overlay(pydub_convert(pydub_load(bass))).overlay(pydub_convert(pydub_load(other))).export(os.path.join(output, f"Instruments.{format}"), format=format)
-
-    for f in [drums, bass, other]:
-        if os.path.exists(f): os.remove(f)
-    
-    logger.info(translations["separator_success_2"])
-    return os.path.join(output, f"Original_Vocals.{format}"), os.path.join(output, f"Instruments.{format}")
-
-def separator_backing(input, output, format, segments_size, overlap, denoise, kara_model, hop_length, batch_size, sample_rate):
-    if not os.path.exists(input): 
-        logger.warning(translations["input_not_valid"])
-        sys.exit(1)
-    
-    if not os.path.exists(output): 
-        logger.warning(translations["output_not_valid"])
-        sys.exit(1)
-    
-    for f in [f"Main_Vocals.{format}", f"Backing_Vocals.{format}"]:
-        if os.path.exists(os.path.join(output, f)): os.remove(os.path.join(output, f))
-
-    model_2 = kara_models.get(kara_model)
-    logger.info(f"{translations['separator_process_backing']}...")
-
-    backing_outputs = separator_main(audio_file=input, model_filename=model_2, output_format=format, output_dir=output, mdx_segment_size=segments_size, mdx_overlap=overlap, mdx_batch_size=batch_size, mdx_hop_length=hop_length, mdx_enable_denoise=denoise, sample_rate=sample_rate)
-    main_output = os.path.join(output, f"Main_Vocals.{format}")
-    backing_output = os.path.join(output, f"Backing_Vocals.{format}")
-
-    for f in backing_outputs:
-        path = os.path.join(output, f)
-        if not os.path.exists(path): logger.error(translations["not_found"].format(name=path))
-
-        if '_(Instrumental)_' in f: os.rename(path, backing_output)
-        elif '_(Vocals)_' in f: os.rename(path, main_output)
-
-    logger.info(translations["separator_process_backing_success"])
-    return main_output, backing_output
-
-def separator_music_mdx(input, output, format, segments_size, overlap, denoise, mdx_model, hop_length, batch_size, sample_rate):
-    if not os.path.exists(input): 
-        logger.warning(translations["input_not_valid"])
-        sys.exit(1)
-    
-    if not os.path.exists(output): 
-        logger.warning(translations["output_not_valid"])
-        sys.exit(1)
-
-    for i in [f"Original_Vocals.{format}", f"Instruments.{format}"]:
-        if os.path.exists(os.path.join(output, i)): os.remove(os.path.join(output, i))
-    
-    model_3 = mdx_models.get(mdx_model)
-    logger.info(f"{translations['separator_process_2']}...")
-
-    output_music = separator_main(audio_file=input, model_filename=model_3, output_format=format, output_dir=output, mdx_segment_size=segments_size, mdx_overlap=overlap, mdx_batch_size=batch_size, mdx_hop_length=hop_length, mdx_enable_denoise=denoise, sample_rate=sample_rate)
-    original_output, instruments_output = os.path.join(output, f"Original_Vocals.{format}"), os.path.join(output, f"Instruments.{format}")
-
-    for f in output_music:
-        path = os.path.join(output, f)
-        if not os.path.exists(path): logger.error(translations["not_found"].format(name=path))
-
-        if '_(Instrumental)_' in f: os.rename(path, instruments_output)
-        elif '_(Vocals)_' in f: os.rename(path, original_output)
-
-    logger.info(translations["separator_process_backing_success"])
-    return original_output, instruments_output
-
-def separator_reverb(output, format, segments_size, overlap, denoise, original, backing_reverb, hop_length, batch_size, sample_rate):
-    if not os.path.exists(output): 
-        logger.warning(translations["output_not_valid"])
-        sys.exit(1)
-    
-    for i in [f"Original_Vocals_Reverb.{format}", f"Main_Vocals_Reverb.{format}", f"Original_Vocals_No_Reverb.{format}", f"Main_Vocals_No_Reverb.{format}"]:
-        if os.path.exists(os.path.join(output, i)): os.remove(os.path.join(output, i))
-
-    dereveb_path = []
-
-    if original: 
-        try:
-            dereveb_path.append(os.path.join(output, [f for f in os.listdir(output) if 'Original_Vocals' in f][0]))
-        except IndexError:
-            logger.warning(translations["not_found_original_vocal"])
-            sys.exit(1)
-        
-    if backing_reverb:
-        try:
-            dereveb_path.append(os.path.join(output, [f for f in os.listdir(output) if 'Main_Vocals' in f][0]))
-        except IndexError:
-            logger.warning(translations["not_found_main_vocal"])
-            sys.exit(1)
-    
-    if backing_reverb:
-        try:
-            dereveb_path.append(os.path.join(output, [f for f in os.listdir(output) if 'Backing_Vocals' in f][0]))
-        except IndexError:
-            logger.warning(translations["not_found_backing_vocal"])
-            sys.exit(1)
-    
-    for path in dereveb_path:
-        if not os.path.exists(path): 
-            logger.warning(translations["not_found"].format(name=path))
-            sys.exit(1)
-        
-        if "Original_Vocals" in path: 
-            reverb_path, no_reverb_path = os.path.join(output, f"Original_Vocals_Reverb.{format}"), os.path.join(output, f"Original_Vocals_No_Reverb.{format}")
-            start_title, end_title = translations["process_original"], translations["process_original_success"]
-        elif "Main_Vocals" in path:
-            reverb_path, no_reverb_path = os.path.join(output, f"Main_Vocals_Reverb.{format}"), os.path.join(output, f"Main_Vocals_No_Reverb.{format}")
-            start_title, end_title = translations["process_main"], translations["process_main_success"]
-        elif "Backing_Vocals" in path:
-            reverb_path, no_reverb_path = os.path.join(output, f"Backing_Vocals_Reverb.{format}"), os.path.join(output, f"Backing_Vocals_No_Reverb.{format}")
-            start_title, end_title = translations["process_backing"], translations["process_backing_success"]
-
-        logger.info(start_title)
-        output_dereveb = separator_main(audio_file=path, model_filename="Reverb_HQ_By_FoxJoy.onnx", output_format=format, output_dir=output, mdx_segment_size=segments_size, mdx_overlap=overlap, mdx_batch_size=batch_size, mdx_hop_length=hop_length, mdx_enable_denoise=denoise, sample_rate=sample_rate)
-
-        for f in output_dereveb:
-            path = os.path.join(output, f)
-            if not os.path.exists(path): logger.error(translations["not_found"].format(name=path))
-
-            if '_(Reverb)_' in f: os.rename(path, reverb_path)
-            elif '_(No Reverb)_' in f: os.rename(path, no_reverb_path)
-
-        logger.info(end_title)
-
-    return (os.path.join(output, f"Original_Vocals_No_Reverb.{format}") if original else None), (os.path.join(output, f"Main_Vocals_No_Reverb.{format}") if backing_reverb else None), (os.path.join(output, f"Backing_Vocals_No_Reverb.{format}") if backing_reverb else None)
-
-def separator_main(audio_file=None, model_filename="UVR-MDX-NET_Main_340.onnx", output_format="wav", output_dir=".", mdx_segment_size=256, mdx_overlap=0.25, mdx_batch_size=1, mdx_hop_length=1024, mdx_enable_denoise=True, demucs_segment_size=256, demucs_shifts=2, demucs_overlap=0.25, sample_rate=44100):
-    try:
-        separator = Separator(logger=logger, log_formatter=file_formatter, log_level=logging.INFO, output_dir=output_dir, output_format=output_format, output_bitrate=None, normalization_threshold=0.9, output_single_stem=None, invert_using_spec=False, sample_rate=sample_rate, mdx_params={"hop_length": mdx_hop_length, "segment_size": mdx_segment_size, "overlap": mdx_overlap, "batch_size": mdx_batch_size, "enable_denoise": mdx_enable_denoise}, demucs_params={"segment_size": demucs_segment_size, "shifts": demucs_shifts, "overlap": demucs_overlap, "segments_enabled": True})
-        separator.load_model(model_filename=model_filename)
-
-        return separator.separate(audio_file)
-    except:
-        logger.debug(translations["default_setting"])
-        separator = Separator(logger=logger, log_formatter=file_formatter, log_level=logging.INFO, output_dir=output_dir, output_format=output_format, output_bitrate=None, normalization_threshold=0.9, output_single_stem=None, invert_using_spec=False, sample_rate=44100, mdx_params={"hop_length": 1024, "segment_size": 256, "overlap": 0.25, "batch_size": 1, "enable_denoise": mdx_enable_denoise}, demucs_params={"segment_size": 128, "shifts": 2, "overlap": 0.25, "segments_enabled": True})
-        separator.load_model(model_filename=model_filename)
-
-        return separator.separate(audio_file)
-
-if __name__ == "__main__": main()

main/inference/train.py

@@ -1,990 +0,0 @@
-import os
-import sys
-import glob
-import json
-import torch
-import hashlib
-import logging
-import argparse
-import datetime
-import warnings
-import logging.handlers
-
-import numpy as np
-import soundfile as sf
-import matplotlib.pyplot as plt
-import torch.distributed as dist
-import torch.utils.data as tdata
-import torch.multiprocessing as mp
-
-from tqdm import tqdm
-from collections import OrderedDict
-from random import randint, shuffle
-from torch.utils.checkpoint import checkpoint
-from torch.cuda.amp import GradScaler, autocast
-from torch.utils.tensorboard import SummaryWriter
-
-from time import time as ttime
-from torch.nn import functional as F
-from distutils.util import strtobool
-from librosa.filters import mel as librosa_mel_fn
-from torch.nn.parallel import DistributedDataParallel as DDP
-from torch.nn.utils.parametrizations import spectral_norm, weight_norm
-
-sys.path.append(os.getcwd())
-
-from main.configs.config import Config
-from main.library.algorithm.residuals import LRELU_SLOPE
-from main.library.algorithm.synthesizers import Synthesizer
-from main.library.algorithm.commons import get_padding, slice_segments, clip_grad_value
-
-MATPLOTLIB_FLAG = False
-main_config = Config()
-translations = main_config.translations
-warnings.filterwarnings("ignore")
-logging.getLogger("torch").setLevel(logging.ERROR)
-
-class HParams:
-    def __init__(self, **kwargs):
-        for k, v in kwargs.items():
-            self[k] = HParams(**v) if isinstance(v, dict) else v
-
-    def keys(self):
-        return self.__dict__.keys()
-    
-    def items(self):
-        return self.__dict__.items()
-    
-    def values(self):
-        return self.__dict__.values()
-    
-    def __len__(self):
-        return len(self.__dict__)
-    
-    def __getitem__(self, key):
-        return self.__dict__[key]
-
-    def __setitem__(self, key, value):
-        self.__dict__[key] = value
-
-    def __contains__(self, key):
-        return key in self.__dict__
-    
-    def __repr__(self):
-        return repr(self.__dict__)
-
-def parse_arguments():
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--model_name", type=str, required=True)
-    parser.add_argument("--rvc_version", type=str, default="v2")
-    parser.add_argument("--save_every_epoch", type=int, required=True)
-    parser.add_argument("--save_only_latest", type=lambda x: bool(strtobool(x)), default=True)
-    parser.add_argument("--save_every_weights", type=lambda x: bool(strtobool(x)), default=True)
-    parser.add_argument("--total_epoch", type=int, default=300)
-    parser.add_argument("--sample_rate", type=int, required=True)
-    parser.add_argument("--batch_size", type=int, default=8)
-    parser.add_argument("--gpu", type=str, default="0")
-    parser.add_argument("--pitch_guidance", type=lambda x: bool(strtobool(x)), default=True)
-    parser.add_argument("--g_pretrained_path", type=str, default="")
-    parser.add_argument("--d_pretrained_path", type=str, default="")
-    parser.add_argument("--overtraining_detector", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--overtraining_threshold", type=int, default=50)
-    parser.add_argument("--cleanup", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--cache_data_in_gpu", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--model_author", type=str)
-    parser.add_argument("--vocoder", type=str, default="Default")
-    parser.add_argument("--checkpointing", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--deterministic", type=lambda x: bool(strtobool(x)), default=False)
-    parser.add_argument("--benchmark", type=lambda x: bool(strtobool(x)), default=False)
-
-    return parser.parse_args()
-
-args = parse_arguments()
-model_name, save_every_epoch, total_epoch, pretrainG, pretrainD, version, gpus, batch_size, sample_rate, pitch_guidance, save_only_latest, save_every_weights, cache_data_in_gpu, overtraining_detector, overtraining_threshold, cleanup, model_author, vocoder, checkpointing = args.model_name, args.save_every_epoch, args.total_epoch, args.g_pretrained_path, args.d_pretrained_path, args.rvc_version, args.gpu, args.batch_size, args.sample_rate, args.pitch_guidance, args.save_only_latest, args.save_every_weights, args.cache_data_in_gpu, args.overtraining_detector, args.overtraining_threshold, args.cleanup, args.model_author, args.vocoder, args.checkpointing
-
-experiment_dir = os.path.join("assets", "logs", model_name)
-training_file_path = os.path.join(experiment_dir, "training_data.json")
-config_save_path = os.path.join(experiment_dir, "config.json")
-torch.backends.cudnn.deterministic = args.deterministic
-torch.backends.cudnn.benchmark = args.benchmark
-lowest_value = {"step": 0, "value": float("inf"), "epoch": 0}
-global_step, last_loss_gen_all, overtrain_save_epoch = 0, 0, 0
-loss_gen_history, smoothed_loss_gen_history, loss_disc_history, smoothed_loss_disc_history = [], [], [], []
-
-with open(config_save_path, "r") as f:
-    config = json.load(f)
-
-config = HParams(**config)
-config.data.training_files = os.path.join(experiment_dir, "filelist.txt")
-logger = logging.getLogger(__name__)
-
-if logger.hasHandlers(): logger.handlers.clear()
-else:  
-    console_handler = logging.StreamHandler()
-    console_handler.setFormatter(logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S"))
-    console_handler.setLevel(logging.INFO)
-    file_handler = logging.handlers.RotatingFileHandler(os.path.join(experiment_dir, "train.log"), maxBytes=5*1024*1024, backupCount=3, encoding='utf-8')
-    file_handler.setFormatter(logging.Formatter(fmt="\n%(asctime)s.%(msecs)03d | %(levelname)s | %(module)s | %(message)s", datefmt="%Y-%m-%d %H:%M:%S"))
-    file_handler.setLevel(logging.DEBUG)
-    logger.addHandler(console_handler)
-    logger.addHandler(file_handler)
-    logger.setLevel(logging.DEBUG)
-
-log_data = {translations['modelname']: model_name, translations["save_every_epoch"]: save_every_epoch, translations["total_e"]: total_epoch, translations["dorg"].format(pretrainG=pretrainG, pretrainD=pretrainD): "", translations['training_version']: version, "Gpu": gpus, translations['batch_size']: batch_size, translations['pretrain_sr']: sample_rate, translations['training_f0']: pitch_guidance, translations['save_only_latest']: save_only_latest, translations['save_every_weights']: save_every_weights, translations['cache_in_gpu']: cache_data_in_gpu, translations['overtraining_detector']: overtraining_detector, translations['threshold']: overtraining_threshold, translations['cleanup_training']: cleanup, translations['memory_efficient_training']: checkpointing}
-if model_author: log_data[translations["model_author"].format(model_author=model_author)] = ""
-if vocoder != "Default": log_data[translations['vocoder']] = vocoder
-
-for key, value in log_data.items():
-    logger.debug(f"{key}: {value}" if value != "" else f"{key} {value}")
-
-def main():
-    global training_file_path, last_loss_gen_all, smoothed_loss_gen_history, loss_gen_history, loss_disc_history, smoothed_loss_disc_history, overtrain_save_epoch, model_author, vocoder, checkpointing, gpus
-    
-    try:
-        os.environ["MASTER_ADDR"] = "localhost"
-        os.environ["MASTER_PORT"] = str(randint(20000, 55555))
-
-        if torch.cuda.is_available():
-            device, gpus = torch.device("cuda"), [int(item) for item in gpus.split("-")]
-            n_gpus = len(gpus)
-        elif torch.backends.mps.is_available():
-            device, gpus = torch.device("mps"), [0]
-            n_gpus = 1
-        else:
-            device, gpus = torch.device("cpu"), [0]
-            n_gpus = 1
-            logger.warning(translations["not_gpu"])
-
-        def start():
-            children = []
-            pid_data = {"process_pids": []}
-
-            with open(config_save_path, "r") as pid_file:
-                try:
-                    pid_data.update(json.load(pid_file))
-                except json.JSONDecodeError:
-                    pass
-
-            with open(config_save_path, "w") as pid_file:
-                for rank, device_id in enumerate(gpus):
-                    subproc = mp.Process(target=run, args=(rank, n_gpus, experiment_dir, pretrainG, pretrainD, pitch_guidance, total_epoch, save_every_weights, config, device, device_id, model_author, vocoder, checkpointing))
-                    children.append(subproc)
-                    subproc.start()
-                    pid_data["process_pids"].append(subproc.pid)
-
-                json.dump(pid_data, pid_file, indent=4)
-
-            for i in range(n_gpus):
-                children[i].join()
-
-        def load_from_json(file_path):
-            if os.path.exists(file_path):
-                with open(file_path, "r") as f:
-                    data = json.load(f)
-                    return (data.get("loss_disc_history", []), data.get("smoothed_loss_disc_history", []), data.get("loss_gen_history", []), data.get("smoothed_loss_gen_history", []))
-            return [], [], [], []
-
-        def continue_overtrain_detector(training_file_path):
-            if overtraining_detector and os.path.exists(training_file_path): (loss_disc_history, smoothed_loss_disc_history, loss_gen_history, smoothed_loss_gen_history) = load_from_json(training_file_path)
-
-        if cleanup:
-            for root, dirs, files in os.walk(experiment_dir, topdown=False):
-                for name in files:
-                    file_path = os.path.join(root, name)
-                    _, file_extension = os.path.splitext(name)
-                    if (file_extension == ".0" or (name.startswith("D_") and file_extension == ".pth") or (name.startswith("G_") and file_extension == ".pth") or (file_extension == ".index")): os.remove(file_path)
-
-                for name in dirs:
-                    if name == "eval":
-                        folder_path = os.path.join(root, name)
-                        for item in os.listdir(folder_path):
-                            item_path = os.path.join(folder_path, item)
-                            if os.path.isfile(item_path): os.remove(item_path)
-                        os.rmdir(folder_path)
-
-        continue_overtrain_detector(training_file_path)
-        start()
-    except Exception as e:
-        logger.error(f"{translations['training_error']} {e}")
-        import traceback
-        logger.debug(traceback.format_exc())
-
-def plot_spectrogram_to_numpy(spectrogram):
-    global MATPLOTLIB_FLAG
-
-    if not MATPLOTLIB_FLAG:
-        plt.switch_backend("Agg")
-        MATPLOTLIB_FLAG = True
-
-    fig, ax = plt.subplots(figsize=(10, 2))
-    plt.colorbar(ax.imshow(spectrogram, aspect="auto", origin="lower", interpolation="none"), ax=ax)
-    plt.xlabel("Frames")
-    plt.ylabel("Channels")
-    plt.tight_layout()
-    fig.canvas.draw()
-    plt.close(fig)
-
-    try:
-        data = np.array(fig.canvas.renderer.buffer_rgba(), dtype=np.uint8).reshape(fig.canvas.get_width_height()[::-1] + (4,))[:, :, :3]
-    except:
-        data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep="").reshape(fig.canvas.get_width_height()[::-1] + (3,))
-
-    return data
-
-def verify_checkpoint_shapes(checkpoint_path, model):
-    checkpoint = torch.load(checkpoint_path, map_location="cpu")
-    checkpoint_state_dict = checkpoint["model"]
-    try:
-        model_state_dict = model.module.load_state_dict(checkpoint_state_dict) if hasattr(model, "module") else model.load_state_dict(checkpoint_state_dict)
-    except RuntimeError:
-        logger.warning(translations["checkpointing_err"])
-        sys.exit(1)
-    else: del checkpoint, checkpoint_state_dict, model_state_dict
-
-def summarize(writer, global_step, scalars={}, histograms={}, images={}, audios={}, audio_sample_rate=22050):
-    for k, v in scalars.items():
-        writer.add_scalar(k, v, global_step)
-
-    for k, v in histograms.items():
-        writer.add_histogram(k, v, global_step)
-
-    for k, v in images.items():
-        writer.add_image(k, v, global_step, dataformats="HWC")
-
-    for k, v in audios.items():
-        writer.add_audio(k, v, global_step, audio_sample_rate)
-
-def load_checkpoint(checkpoint_path, model, optimizer=None, load_opt=1):
-    assert os.path.isfile(checkpoint_path), translations["not_found_checkpoint"].format(checkpoint_path=checkpoint_path)
-    checkpoint_dict = replace_keys_in_dict(replace_keys_in_dict(torch.load(checkpoint_path, map_location="cpu"), ".weight_v", ".parametrizations.weight.original1"), ".weight_g", ".parametrizations.weight.original0")
-    new_state_dict = {k: checkpoint_dict["model"].get(k, v) for k, v in (model.module.state_dict() if hasattr(model, "module") else model.state_dict()).items()}
-
-    if hasattr(model, "module"): model.module.load_state_dict(new_state_dict, strict=False)
-    else: model.load_state_dict(new_state_dict, strict=False)
-
-    if optimizer and load_opt == 1: optimizer.load_state_dict(checkpoint_dict.get("optimizer", {}))
-    logger.debug(translations["save_checkpoint"].format(checkpoint_path=checkpoint_path, checkpoint_dict=checkpoint_dict['iteration']))
-    return (model, optimizer, checkpoint_dict.get("learning_rate", 0), checkpoint_dict["iteration"])
-
-def save_checkpoint(model, optimizer, learning_rate, iteration, checkpoint_path):
-    state_dict = (model.module.state_dict() if hasattr(model, "module") else model.state_dict())
-    torch.save(replace_keys_in_dict(replace_keys_in_dict({"model": state_dict, "iteration": iteration, "optimizer": optimizer.state_dict(), "learning_rate": learning_rate}, ".parametrizations.weight.original1", ".weight_v"), ".parametrizations.weight.original0", ".weight_g"), checkpoint_path)
-    logger.info(translations["save_model"].format(checkpoint_path=checkpoint_path, iteration=iteration))
-
-def latest_checkpoint_path(dir_path, regex="G_*.pth"):
-    checkpoints = sorted(glob.glob(os.path.join(dir_path, regex)), key=lambda f: int("".join(filter(str.isdigit, f))))
-    return checkpoints[-1] if checkpoints else None
-
-def load_wav_to_torch(full_path):
-    data, sample_rate = sf.read(full_path, dtype=np.float32)
-    return torch.FloatTensor(data.astype(np.float32)), sample_rate
-
-def load_filepaths_and_text(filename, split="|"):
-    with open(filename, encoding="utf-8") as f:
-        return [line.strip().split(split) for line in f]
-
-def feature_loss(fmap_r, fmap_g):
-    loss = 0
-    for dr, dg in zip(fmap_r, fmap_g):
-        for rl, gl in zip(dr, dg):
-            loss += torch.mean(torch.abs(rl.float().detach() - gl.float()))
-    return loss * 2
-
-def discriminator_loss(disc_real_outputs, disc_generated_outputs):
-    loss = 0
-    r_losses, g_losses = [], []
-
-    for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
-        dr = dr.float()
-        dg = dg.float()
-        r_loss = torch.mean((1 - dr) ** 2)
-        g_loss = torch.mean(dg**2)
-        loss += r_loss + g_loss
-        r_losses.append(r_loss.item())
-        g_losses.append(g_loss.item())
-    return loss, r_losses, g_losses
-
-def generator_loss(disc_outputs):
-    loss = 0
-    gen_losses = []
-
-    for dg in disc_outputs:
-        l = torch.mean((1 - dg.float()) ** 2)
-        gen_losses.append(l)
-        loss += l
-    return loss, gen_losses
-
-def kl_loss(z_p, logs_q, m_p, logs_p, z_mask):
-    z_p = z_p.float()
-    logs_q = logs_q.float()
-    m_p = m_p.float()
-    logs_p = logs_p.float()
-    z_mask = z_mask.float()
-    kl = logs_p - logs_q - 0.5
-    kl += 0.5 * ((z_p - m_p) ** 2) * torch.exp(-2.0 * logs_p)
-    return torch.sum(kl * z_mask) / torch.sum(z_mask)
-
-class TextAudioLoaderMultiNSFsid(tdata.Dataset):
-    def __init__(self, hparams):
-        self.audiopaths_and_text = load_filepaths_and_text(hparams.training_files)
-        self.max_wav_value = hparams.max_wav_value
-        self.sample_rate = hparams.sample_rate
-        self.filter_length = hparams.filter_length
-        self.hop_length = hparams.hop_length
-        self.win_length = hparams.win_length
-        self.sample_rate = hparams.sample_rate
-        self.min_text_len = getattr(hparams, "min_text_len", 1)
-        self.max_text_len = getattr(hparams, "max_text_len", 5000)
-        self._filter()
-
-    def _filter(self):
-        audiopaths_and_text_new, lengths = [], []
-        for audiopath, text, pitch, pitchf, dv in self.audiopaths_and_text:
-            if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
-                audiopaths_and_text_new.append([audiopath, text, pitch, pitchf, dv])
-                lengths.append(os.path.getsize(audiopath) // (3 * self.hop_length))
-
-        self.audiopaths_and_text = audiopaths_and_text_new
-        self.lengths = lengths
-
-    def get_sid(self, sid):
-        try:
-            sid = torch.LongTensor([int(sid)])
-        except ValueError as e:
-            logger.error(translations["sid_error"].format(sid=sid, e=e))
-            sid = torch.LongTensor([0])
-        return sid
-
-    def get_audio_text_pair(self, audiopath_and_text):
-        phone, pitch, pitchf = self.get_labels(audiopath_and_text[1], audiopath_and_text[2], audiopath_and_text[3])
-        spec, wav = self.get_audio(audiopath_and_text[0])
-        dv = self.get_sid(audiopath_and_text[4])
-        len_phone = phone.size()[0]
-        len_spec = spec.size()[-1]
-
-        if len_phone != len_spec:
-            len_min = min(len_phone, len_spec)
-            len_wav = len_min * self.hop_length
-            spec, wav, phone = spec[:, :len_min], wav[:, :len_wav], phone[:len_min, :]
-            pitch, pitchf = pitch[:len_min], pitchf[:len_min]
-        return (spec, wav, phone, pitch, pitchf, dv)
-
-    def get_labels(self, phone, pitch, pitchf):
-        phone = np.repeat(np.load(phone), 2, axis=0)
-        n_num = min(phone.shape[0], 900)
-        return torch.FloatTensor(phone[:n_num, :]), torch.LongTensor(np.load(pitch)[:n_num]), torch.FloatTensor(np.load(pitchf)[:n_num])
-
-    def get_audio(self, filename):
-        audio, sample_rate = load_wav_to_torch(filename)
-        if sample_rate != self.sample_rate: raise ValueError(translations["sr_does_not_match"].format(sample_rate=sample_rate, sample_rate2=self.sample_rate))
-        audio_norm = audio.unsqueeze(0)
-        spec_filename = filename.replace(".wav", ".spec.pt")
-
-        if os.path.exists(spec_filename):
-            try:
-                spec = torch.load(spec_filename)
-            except Exception as e:
-                logger.error(translations["spec_error"].format(spec_filename=spec_filename, e=e))
-                spec = torch.squeeze(spectrogram_torch(audio_norm, self.filter_length, self.hop_length, self.win_length, center=False), 0)
-                torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
-        else: 
-            spec = torch.squeeze(spectrogram_torch(audio_norm, self.filter_length, self.hop_length, self.win_length, center=False), 0)
-            torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
-        return spec, audio_norm
-
-    def __getitem__(self, index):
-        return self.get_audio_text_pair(self.audiopaths_and_text[index])
-
-    def __len__(self):
-        return len(self.audiopaths_and_text)
-
-class TextAudioCollateMultiNSFsid:
-    def __init__(self, return_ids=False):
-        self.return_ids = return_ids
-
-    def __call__(self, batch):
-        _, ids_sorted_decreasing = torch.sort(torch.LongTensor([x[0].size(1) for x in batch]), dim=0, descending=True)
-        spec_lengths, wave_lengths = torch.LongTensor(len(batch)), torch.LongTensor(len(batch))
-        spec_padded, wave_padded = torch.FloatTensor(len(batch), batch[0][0].size(0), max([x[0].size(1) for x in batch])), torch.FloatTensor(len(batch), 1, max([x[1].size(1) for x in batch]))
-        spec_padded.zero_()
-        wave_padded.zero_()
-        max_phone_len = max([x[2].size(0) for x in batch])
-        phone_lengths, phone_padded = torch.LongTensor(len(batch)), torch.FloatTensor(len(batch), max_phone_len, batch[0][2].shape[1])
-        pitch_padded, pitchf_padded = torch.LongTensor(len(batch), max_phone_len), torch.FloatTensor(len(batch), max_phone_len)
-        phone_padded.zero_()
-        pitch_padded.zero_()
-        pitchf_padded.zero_()
-        sid = torch.LongTensor(len(batch))
-
-        for i in range(len(ids_sorted_decreasing)):
-            row = batch[ids_sorted_decreasing[i]]
-            spec = row[0]
-            spec_padded[i, :, : spec.size(1)] = spec
-            spec_lengths[i] = spec.size(1)
-            wave = row[1]
-            wave_padded[i, :, : wave.size(1)] = wave
-            wave_lengths[i] = wave.size(1)
-            phone = row[2]
-            phone_padded[i, : phone.size(0), :] = phone
-            phone_lengths[i] = phone.size(0)
-            pitch = row[3]
-            pitch_padded[i, : pitch.size(0)] = pitch
-            pitchf = row[4]
-            pitchf_padded[i, : pitchf.size(0)] = pitchf
-            sid[i] = row[5]
-        return (phone_padded, phone_lengths, pitch_padded, pitchf_padded, spec_padded, spec_lengths, wave_padded, wave_lengths, sid)
-
-class TextAudioLoader(tdata.Dataset):
-    def __init__(self, hparams):
-        self.audiopaths_and_text = load_filepaths_and_text(hparams.training_files)
-        self.max_wav_value = hparams.max_wav_value
-        self.sample_rate = hparams.sample_rate
-        self.filter_length = hparams.filter_length
-        self.hop_length = hparams.hop_length
-        self.win_length = hparams.win_length
-        self.sample_rate = hparams.sample_rate
-        self.min_text_len = getattr(hparams, "min_text_len", 1)
-        self.max_text_len = getattr(hparams, "max_text_len", 5000)
-        self._filter()
-
-    def _filter(self):
-        audiopaths_and_text_new, lengths = [], []
-        for entry in self.audiopaths_and_text:
-            if len(entry) >= 3:
-                audiopath, text, dv = entry[:3]
-                if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
-                    audiopaths_and_text_new.append([audiopath, text, dv])
-                    lengths.append(os.path.getsize(audiopath) // (3 * self.hop_length))
-
-        self.audiopaths_and_text = audiopaths_and_text_new
-        self.lengths = lengths
-
-    def get_sid(self, sid):
-        try:
-            sid = torch.LongTensor([int(sid)])
-        except ValueError as e:
-            logger.error(translations["sid_error"].format(sid=sid, e=e))
-            sid = torch.LongTensor([0])
-        return sid
-
-    def get_audio_text_pair(self, audiopath_and_text):
-        phone = self.get_labels(audiopath_and_text[1])
-        spec, wav = self.get_audio(audiopath_and_text[0])
-        dv = self.get_sid(audiopath_and_text[2])
-        len_phone = phone.size()[0]
-        len_spec = spec.size()[-1]
-
-        if len_phone != len_spec:
-            len_min = min(len_phone, len_spec)
-            len_wav = len_min * self.hop_length
-            spec = spec[:, :len_min]
-            wav = wav[:, :len_wav]
-            phone = phone[:len_min, :]
-        return (spec, wav, phone, dv)
-
-    def get_labels(self, phone):
-        phone = np.repeat(np.load(phone), 2, axis=0)
-        return torch.FloatTensor(phone[:min(phone.shape[0], 900), :])
-
-    def get_audio(self, filename):
-        audio, sample_rate = load_wav_to_torch(filename)
-        if sample_rate != self.sample_rate: raise ValueError(translations["sr_does_not_match"].format(sample_rate=sample_rate, sample_rate2=self.sample_rate))
-        audio_norm = audio.unsqueeze(0)
-        spec_filename = filename.replace(".wav", ".spec.pt")
-
-        if os.path.exists(spec_filename):
-            try:
-                spec = torch.load(spec_filename)
-            except Exception as e:
-                logger.error(translations["spec_error"].format(spec_filename=spec_filename, e=e))
-                spec = torch.squeeze(spectrogram_torch(audio_norm, self.filter_length, self.hop_length, self.win_length, center=False), 0)
-                torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
-        else:
-            spec = torch.squeeze(spectrogram_torch(audio_norm, self.filter_length, self.hop_length, self.win_length, center=False), 0)
-            torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
-        return spec, audio_norm
-
-    def __getitem__(self, index):
-        return self.get_audio_text_pair(self.audiopaths_and_text[index])
-
-    def __len__(self):
-        return len(self.audiopaths_and_text)
-
-class TextAudioCollate:
-    def __init__(self, return_ids=False):
-        self.return_ids = return_ids
-
-    def __call__(self, batch):
-        _, ids_sorted_decreasing = torch.sort(torch.LongTensor([x[0].size(1) for x in batch]), dim=0, descending=True)
-        spec_lengths, wave_lengths = torch.LongTensor(len(batch)), torch.LongTensor(len(batch))
-        spec_padded, wave_padded = torch.FloatTensor(len(batch), batch[0][0].size(0), max([x[0].size(1) for x in batch])), torch.FloatTensor(len(batch), 1, max([x[1].size(1) for x in batch]))
-        spec_padded.zero_()
-        wave_padded.zero_()
-        max_phone_len = max([x[2].size(0) for x in batch])
-        phone_lengths, phone_padded = torch.LongTensor(len(batch)), torch.FloatTensor(len(batch), max_phone_len, batch[0][2].shape[1])
-        phone_padded.zero_()
-        sid = torch.LongTensor(len(batch))
-        for i in range(len(ids_sorted_decreasing)):
-            row = batch[ids_sorted_decreasing[i]]
-            spec = row[0]
-            spec_padded[i, :, : spec.size(1)] = spec
-            spec_lengths[i] = spec.size(1)
-            wave = row[1]
-            wave_padded[i, :, : wave.size(1)] = wave
-            wave_lengths[i] = wave.size(1)
-            phone = row[2]
-            phone_padded[i, : phone.size(0), :] = phone
-            phone_lengths[i] = phone.size(0)
-            sid[i] = row[3]
-        return (phone_padded, phone_lengths, spec_padded, spec_lengths, wave_padded, wave_lengths, sid)
-
-class DistributedBucketSampler(tdata.distributed.DistributedSampler):
-    def __init__(self, dataset, batch_size, boundaries, num_replicas=None, rank=None, shuffle=True):
-        super().__init__(dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle)
-        self.lengths = dataset.lengths
-        self.batch_size = batch_size
-        self.boundaries = boundaries
-        self.buckets, self.num_samples_per_bucket = self._create_buckets()
-        self.total_size = sum(self.num_samples_per_bucket)
-        self.num_samples = self.total_size // self.num_replicas
-
-    def _create_buckets(self):
-        buckets = [[] for _ in range(len(self.boundaries) - 1)]
-        for i in range(len(self.lengths)):
-            idx_bucket = self._bisect(self.lengths[i])
-            if idx_bucket != -1: buckets[idx_bucket].append(i)
-
-        for i in range(len(buckets) - 1, -1, -1):  
-            if len(buckets[i]) == 0:
-                buckets.pop(i)
-                self.boundaries.pop(i + 1)
-
-        num_samples_per_bucket = []
-        for i in range(len(buckets)):
-            len_bucket = len(buckets[i])
-            total_batch_size = self.num_replicas * self.batch_size
-            num_samples_per_bucket.append(len_bucket + ((total_batch_size - (len_bucket % total_batch_size)) % total_batch_size))
-        return buckets, num_samples_per_bucket
-
-    def __iter__(self):
-        g = torch.Generator()
-        g.manual_seed(self.epoch)
-        indices, batches = [], []
-        if self.shuffle:
-            for bucket in self.buckets:
-                indices.append(torch.randperm(len(bucket), generator=g).tolist())
-        else:
-            for bucket in self.buckets:
-                indices.append(list(range(len(bucket))))
-
-        for i in range(len(self.buckets)):
-            bucket = self.buckets[i]
-            len_bucket = len(bucket)
-            ids_bucket = indices[i]
-            rem = self.num_samples_per_bucket[i] - len_bucket
-            ids_bucket = (ids_bucket + ids_bucket * (rem // len_bucket) + ids_bucket[: (rem % len_bucket)])[self.rank :: self.num_replicas]
-
-            for j in range(len(ids_bucket) // self.batch_size):
-                batches.append([bucket[idx] for idx in ids_bucket[j * self.batch_size : (j + 1) * self.batch_size]])
-
-        if self.shuffle: batches = [batches[i] for i in torch.randperm(len(batches), generator=g).tolist()]
-        self.batches = batches
-        assert len(self.batches) * self.batch_size == self.num_samples
-        return iter(self.batches)
-
-    def _bisect(self, x, lo=0, hi=None):
-        if hi is None: hi = len(self.boundaries) - 1
-
-        if hi > lo:
-            mid = (hi + lo) // 2
-            if self.boundaries[mid] < x and x <= self.boundaries[mid + 1]: return mid
-            elif x <= self.boundaries[mid]: return self._bisect(x, lo, mid)
-            else: return self._bisect(x, mid + 1, hi)
-        else: return -1
-
-    def __len__(self):
-        return self.num_samples // self.batch_size
-
-class MultiPeriodDiscriminator(torch.nn.Module):
-    def __init__(self, version, use_spectral_norm=False, checkpointing=False):
-        super(MultiPeriodDiscriminator, self).__init__()
-        self.checkpointing = checkpointing
-        periods = ([2, 3, 5, 7, 11, 17] if version == "v1" else [2, 3, 5, 7, 11, 17, 23, 37])
-        self.discriminators = torch.nn.ModuleList([DiscriminatorS(use_spectral_norm=use_spectral_norm, checkpointing=checkpointing)] + [DiscriminatorP(p, use_spectral_norm=use_spectral_norm, checkpointing=checkpointing) for p in periods])
-
-    def forward(self, y, y_hat):
-        y_d_rs, y_d_gs, fmap_rs, fmap_gs = [], [], [], []
-        for d in self.discriminators:
-            if self.training and self.checkpointing:
-                def forward_discriminator(d, y, y_hat):
-                    y_d_r, fmap_r = d(y)
-                    y_d_g, fmap_g = d(y_hat)
-                    return y_d_r, fmap_r, y_d_g, fmap_g
-                y_d_r, fmap_r, y_d_g, fmap_g = checkpoint(forward_discriminator, d, y, y_hat, use_reentrant=False)
-            else:
-                y_d_r, fmap_r = d(y)
-                y_d_g, fmap_g = d(y_hat)
-
-            y_d_rs.append(y_d_r); fmap_rs.append(fmap_r)
-            y_d_gs.append(y_d_g); fmap_gs.append(fmap_g)
-        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
-
-class DiscriminatorS(torch.nn.Module):
-    def __init__(self, use_spectral_norm=False, checkpointing=False):
-        super(DiscriminatorS, self).__init__()
-        self.checkpointing = checkpointing
-        norm_f = spectral_norm if use_spectral_norm else weight_norm
-        self.convs = torch.nn.ModuleList([norm_f(torch.nn.Conv1d(1, 16, 15, 1, padding=7)), norm_f(torch.nn.Conv1d(16, 64, 41, 4, groups=4, padding=20)), norm_f(torch.nn.Conv1d(64, 256, 41, 4, groups=16, padding=20)), norm_f(torch.nn.Conv1d(256, 1024, 41, 4, groups=64, padding=20)), norm_f(torch.nn.Conv1d(1024, 1024, 41, 4, groups=256, padding=20)), norm_f(torch.nn.Conv1d(1024, 1024, 5, 1, padding=2))])
-        self.conv_post = norm_f(torch.nn.Conv1d(1024, 1, 3, 1, padding=1))
-        self.lrelu = torch.nn.LeakyReLU(LRELU_SLOPE)
-
-    def forward(self, x):
-        fmap = []
-        for conv in self.convs:
-            x = checkpoint(self.lrelu, checkpoint(conv, x, use_reentrant = False), use_reentrant = False) if self.training and self.checkpointing else self.lrelu(conv(x))
-            fmap.append(x)
-
-        x = self.conv_post(x)
-        fmap.append(x)
-        return torch.flatten(x, 1, -1), fmap
-
-class DiscriminatorP(torch.nn.Module):
-    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False, checkpointing=False):
-        super(DiscriminatorP, self).__init__()
-        self.period = period
-        self.checkpointing = checkpointing
-        norm_f = spectral_norm if use_spectral_norm else weight_norm
-        self.convs = torch.nn.ModuleList([norm_f(torch.nn.Conv2d(in_ch, out_ch, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))) for in_ch, out_ch in zip([1, 32, 128, 512, 1024], [32, 128, 512, 1024, 1024])])
-        self.conv_post = norm_f(torch.nn.Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
-        self.lrelu = torch.nn.LeakyReLU(LRELU_SLOPE)
-
-    def forward(self, x):
-        fmap = []
-        b, c, t = x.shape
-
-        if t % self.period != 0: x = F.pad(x, (0, (self.period - (t % self.period))), "reflect")
-        x = x.view(b, c, -1, self.period)
-
-        for conv in self.convs:
-            x = checkpoint(self.lrelu, checkpoint(conv, x, use_reentrant = False), use_reentrant = False) if self.training and self.checkpointing else self.lrelu(conv(x))
-            fmap.append(x)
-
-        x = self.conv_post(x)
-        fmap.append(x)
-        return torch.flatten(x, 1, -1), fmap
-
-class EpochRecorder:
-    def __init__(self):
-        self.last_time = ttime()
-
-    def record(self):
-        now_time = ttime()
-        elapsed_time = now_time - self.last_time
-        self.last_time = now_time
-        return translations["time_or_speed_training"].format(current_time=datetime.datetime.now().strftime("%H:%M:%S"), elapsed_time_str=str(datetime.timedelta(seconds=int(round(elapsed_time, 1)))))
-    
-def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
-    return torch.log(torch.clamp(x, min=clip_val) * C)
-
-def dynamic_range_decompression_torch(x, C=1):
-    return torch.exp(x) / C
-
-def spectral_normalize_torch(magnitudes):
-    return dynamic_range_compression_torch(magnitudes)
-
-def spectral_de_normalize_torch(magnitudes):
-    return dynamic_range_decompression_torch(magnitudes)
-
-mel_basis, hann_window = {}, {}
-
-def spectrogram_torch(y, n_fft, hop_size, win_size, center=False):
-    global hann_window
-
-    wnsize_dtype_device = str(win_size) + "_" + str(y.dtype) + "_" + str(y.device)
-    if wnsize_dtype_device not in hann_window: hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
-    spec = torch.stft(F.pad(y.unsqueeze(1), (int((n_fft - hop_size) / 2), int((n_fft - hop_size) / 2)), mode="reflect").squeeze(1), n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device], center=center, pad_mode="reflect", normalized=False, onesided=True, return_complex=True)
-    return torch.sqrt(spec.real.pow(2) + spec.imag.pow(2) + 1e-6)
-
-def spec_to_mel_torch(spec, n_fft, num_mels, sample_rate, fmin, fmax):
-    global mel_basis
-
-    fmax_dtype_device = str(fmax) + "_" + str(spec.dtype) + "_" + str(spec.device)
-    if fmax_dtype_device not in mel_basis: mel_basis[fmax_dtype_device] = torch.from_numpy(librosa_mel_fn(sr=sample_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax)).to(dtype=spec.dtype, device=spec.device)
-    return spectral_normalize_torch(torch.matmul(mel_basis[fmax_dtype_device], spec))
-
-def mel_spectrogram_torch(y, n_fft, num_mels, sample_rate, hop_size, win_size, fmin, fmax, center=False):
-    return spec_to_mel_torch(spectrogram_torch(y, n_fft, hop_size, win_size, center), n_fft, num_mels, sample_rate, fmin, fmax)
-
-def replace_keys_in_dict(d, old_key_part, new_key_part):
-    updated_dict = OrderedDict() if isinstance(d, OrderedDict) else {}
-    for key, value in d.items():
-        updated_dict[(key.replace(old_key_part, new_key_part) if isinstance(key, str) else key)] = (replace_keys_in_dict(value, old_key_part, new_key_part) if isinstance(value, dict) else value)
-    return updated_dict
-
-def extract_model(ckpt, sr, pitch_guidance, name, model_path, epoch, step, version, hps, model_author, vocoder):
-    try:
-        logger.info(translations["savemodel"].format(model_dir=model_path, epoch=epoch, step=step))
-        os.makedirs(os.path.dirname(model_path), exist_ok=True)
-
-        opt = OrderedDict(weight={key: value.half() for key, value in ckpt.items() if "enc_q" not in key})
-        opt["config"] = [hps.data.filter_length // 2 + 1, 32, hps.model.inter_channels, hps.model.hidden_channels, hps.model.filter_channels, hps.model.n_heads, hps.model.n_layers, hps.model.kernel_size, hps.model.p_dropout, hps.model.resblock, hps.model.resblock_kernel_sizes, hps.model.resblock_dilation_sizes, hps.model.upsample_rates, hps.model.upsample_initial_channel, hps.model.upsample_kernel_sizes, hps.model.spk_embed_dim, hps.model.gin_channels, hps.data.sample_rate]
-        opt["epoch"] = f"{epoch}epoch"
-        opt["step"] = step
-        opt["sr"] = sr
-        opt["f0"] = int(pitch_guidance)
-        opt["version"] = version
-        opt["creation_date"] = datetime.datetime.now().isoformat()
-        opt["model_hash"] = hashlib.sha256(f"{str(ckpt)} {epoch} {step} {datetime.datetime.now().isoformat()}".encode()).hexdigest()
-        opt["model_name"] = name
-        opt["author"] = model_author
-        opt["vocoder"] = vocoder
-
-        torch.save(replace_keys_in_dict(replace_keys_in_dict(opt, ".parametrizations.weight.original1", ".weight_v"), ".parametrizations.weight.original0", ".weight_g"), model_path)
-    except Exception as e:
-        logger.error(f"{translations['extract_model_error']}: {e}")
-
-def run(rank, n_gpus, experiment_dir, pretrainG, pretrainD, pitch_guidance, custom_total_epoch, custom_save_every_weights, config, device, device_id, model_author, vocoder, checkpointing):
-    global global_step
-
-    if rank == 0: writer_eval = SummaryWriter(log_dir=os.path.join(experiment_dir, "eval"))
-    else: writer_eval = None
-
-    try:
-        dist.init_process_group(backend=("gloo" if sys.platform == "win32" or device.type != "cuda" else "nccl"), init_method="env://", world_size=n_gpus, rank=rank)
-    except:
-        dist.init_process_group(backend=("gloo" if sys.platform == "win32" or device.type != "cuda" else "nccl"), init_method="env://?use_libuv=False", world_size=n_gpus, rank=rank)
-
-    torch.manual_seed(config.train.seed)
-    if torch.cuda.is_available(): torch.cuda.set_device(device_id)
-
-    train_dataset = TextAudioLoaderMultiNSFsid(config.data) if pitch_guidance else TextAudioLoader(config.data)
-    train_loader = tdata.DataLoader(train_dataset, num_workers=4, shuffle=False, pin_memory=True, collate_fn=TextAudioCollateMultiNSFsid() if pitch_guidance else TextAudioCollate(), batch_sampler=DistributedBucketSampler(train_dataset, batch_size * n_gpus, [100, 200, 300, 400, 500, 600, 700, 800, 900], num_replicas=n_gpus, rank=rank, shuffle=True), persistent_workers=True, prefetch_factor=8)
-
-    net_g, net_d = Synthesizer(config.data.filter_length // 2 + 1, config.train.segment_size // config.data.hop_length, **config.model, use_f0=pitch_guidance, sr=sample_rate, vocoder=vocoder, checkpointing=checkpointing), MultiPeriodDiscriminator(version, config.model.use_spectral_norm, checkpointing=checkpointing)
-    net_g, net_d = (net_g.cuda(device_id), net_d.cuda(device_id)) if torch.cuda.is_available() else (net_g.to(device), net_d.to(device))
-
-    optim_g, optim_d = torch.optim.AdamW(net_g.parameters(), config.train.learning_rate, betas=config.train.betas, eps=config.train.eps), torch.optim.AdamW(net_d.parameters(), config.train.learning_rate, betas=config.train.betas, eps=config.train.eps)
-    net_g, net_d = (DDP(net_g, device_ids=[device_id]), DDP(net_d, device_ids=[device_id])) if torch.cuda.is_available() else (DDP(net_g), DDP(net_d))
-
-    try:
-        logger.info(translations["start_training"])
-        _, _, _, epoch_str = load_checkpoint((os.path.join(experiment_dir, "D_latest.pth") if save_only_latest else latest_checkpoint_path(experiment_dir, "D_*.pth")), net_d, optim_d)
-        _, _, _, epoch_str = load_checkpoint((os.path.join(experiment_dir, "G_latest.pth") if save_only_latest else latest_checkpoint_path(experiment_dir, "G_*.pth")), net_g, optim_g)
-        epoch_str += 1
-        global_step = (epoch_str - 1) * len(train_loader)
-    except:
-        epoch_str, global_step = 1, 0
-    
-        if pretrainG != "" and pretrainG != "None":
-            if rank == 0:
-                verify_checkpoint_shapes(pretrainG, net_g)
-                logger.info(translations["import_pretrain"].format(dg="G", pretrain=pretrainG))
-
-            if hasattr(net_g, "module"): net_g.module.load_state_dict(torch.load(pretrainG, map_location="cpu")["model"])
-            else: net_g.load_state_dict(torch.load(pretrainG, map_location="cpu")["model"])
-        else: logger.warning(translations["not_using_pretrain"].format(dg="G"))
-
-        if pretrainD != "" and pretrainD != "None":
-            if rank == 0:
-                verify_checkpoint_shapes(pretrainD, net_d)
-                logger.info(translations["import_pretrain"].format(dg="D", pretrain=pretrainD))
-
-            if hasattr(net_d, "module"): net_d.module.load_state_dict(torch.load(pretrainD, map_location="cpu")["model"])
-            else: net_d.load_state_dict(torch.load(pretrainD, map_location="cpu")["model"])
-        else: logger.warning(translations["not_using_pretrain"].format(dg="D"))
-
-    scheduler_g, scheduler_d = torch.optim.lr_scheduler.ExponentialLR(optim_g, gamma=config.train.lr_decay, last_epoch=epoch_str - 2), torch.optim.lr_scheduler.ExponentialLR(optim_d, gamma=config.train.lr_decay, last_epoch=epoch_str - 2)
-    optim_d.step(); optim_g.step()
-
-    scaler = GradScaler(enabled=main_config.is_half and device.type == "cuda")
-    cache = []
-
-    for info in train_loader:
-        phone, phone_lengths, pitch, pitchf, _, _, _, _, sid = info
-        reference = (phone.cuda(device_id, non_blocking=True), phone_lengths.cuda(device_id, non_blocking=True), (pitch.cuda(device_id, non_blocking=True) if pitch_guidance else None), (pitchf.cuda(device_id, non_blocking=True) if pitch_guidance else None), sid.cuda(device_id, non_blocking=True)) if device.type == "cuda" else (phone.to(device), phone_lengths.to(device), (pitch.to(device) if pitch_guidance else None), (pitchf.to(device) if pitch_guidance else None), sid.to(device))
-        break
-
-    for epoch in range(epoch_str, total_epoch + 1):
-        train_and_evaluate(rank, epoch, config, [net_g, net_d], [optim_g, optim_d], scaler, train_loader, writer_eval, cache, custom_save_every_weights, custom_total_epoch, device, device_id, reference, model_author, vocoder) 
-        scheduler_g.step(); scheduler_d.step()
-
-def train_and_evaluate(rank, epoch, hps, nets, optims, scaler, train_loader, writer, cache, custom_save_every_weights, custom_total_epoch, device, device_id, reference, model_author, vocoder):
-    global global_step, lowest_value, loss_disc, consecutive_increases_gen, consecutive_increases_disc
-
-    if epoch == 1:
-        lowest_value = {"step": 0, "value": float("inf"), "epoch": 0}
-        last_loss_gen_all, consecutive_increases_gen, consecutive_increases_disc = 0.0, 0, 0
-
-    net_g, net_d = nets
-    optim_g, optim_d = optims
-    train_loader.batch_sampler.set_epoch(epoch)
-
-    net_g.train(); net_d.train()
-
-    if device.type == "cuda" and cache_data_in_gpu:
-        data_iterator = cache
-        if cache == []:
-            for batch_idx, info in enumerate(train_loader):
-                cache.append((batch_idx, [tensor.cuda(device_id, non_blocking=True) for tensor in info]))
-        else: shuffle(cache)
-    else: data_iterator = enumerate(train_loader)
-
-    epoch_recorder = EpochRecorder()
-    with tqdm(total=len(train_loader), leave=False) as pbar:
-        for batch_idx, info in data_iterator:
-            if device.type == "cuda" and not cache_data_in_gpu: info = [tensor.cuda(device_id, non_blocking=True) for tensor in info]
-            elif device.type != "cuda": info = [tensor.to(device) for tensor in info]
-
-            phone, phone_lengths, pitch, pitchf, spec, spec_lengths, wave, _, sid = info
-            pitch = pitch if pitch_guidance else None
-            pitchf = pitchf if pitch_guidance else None
-
-            with autocast(enabled=main_config.is_half and device.type == "cuda"):
-                y_hat, ids_slice, _, z_mask, (_, z_p, m_p, logs_p, _, logs_q) = net_g(phone, phone_lengths, pitch, pitchf, spec, spec_lengths, sid)
-                mel = spec_to_mel_torch(spec, config.data.filter_length, config.data.n_mel_channels, config.data.sample_rate, config.data.mel_fmin, config.data.mel_fmax)
-                y_mel = slice_segments(mel, ids_slice, config.train.segment_size // config.data.hop_length, dim=3)
-
-                with autocast(enabled=main_config.is_half and device.type == "cuda"):
-                    y_hat_mel = mel_spectrogram_torch(y_hat.float().squeeze(1), config.data.filter_length, config.data.n_mel_channels, config.data.sample_rate, config.data.hop_length, config.data.win_length, config.data.mel_fmin, config.data.mel_fmax)
-
-                wave = slice_segments(wave, ids_slice * config.data.hop_length, config.train.segment_size, dim=3)
-                y_d_hat_r, y_d_hat_g, _, _ = net_d(wave, y_hat.detach())
-
-                with autocast(enabled=main_config.is_half and device.type == "cuda"):
-                    loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(y_d_hat_r, y_d_hat_g)
-
-            optim_d.zero_grad()
-            scaler.scale(loss_disc).backward()
-            scaler.unscale_(optim_d)
-            grad_norm_d = clip_grad_value(net_d.parameters(), None)
-            scaler.step(optim_d)
-
-            with autocast(enabled=main_config.is_half and device.type == "cuda"):
-                y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(wave, y_hat)
-                with autocast(enabled=main_config.is_half and device.type == "cuda"):
-                    loss_mel = F.l1_loss(y_mel, y_hat_mel) * config.train.c_mel
-                    loss_kl = (kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * config.train.c_kl)
-                    loss_fm = feature_loss(fmap_r, fmap_g)
-                    loss_gen, losses_gen = generator_loss(y_d_hat_g)
-                    loss_gen_all = loss_gen + loss_fm + loss_mel + loss_kl
-                    if loss_gen_all < lowest_value["value"]:
-                        lowest_value["value"] = loss_gen_all
-                        lowest_value["step"] = global_step
-                        lowest_value["epoch"] = epoch
-                        if epoch > lowest_value["epoch"]: logger.warning(translations["training_warning"])
-
-            optim_g.zero_grad()
-            scaler.scale(loss_gen_all).backward()
-            scaler.unscale_(optim_g)
-            grad_norm_g = clip_grad_value(net_g.parameters(), None)
-            scaler.step(optim_g)
-            scaler.update()
-
-            if rank == 0 and global_step % config.train.log_interval == 0:
-                if loss_mel > 75: loss_mel = 75
-                if loss_kl > 9: loss_kl = 9
-
-                scalar_dict = {"loss/g/total": loss_gen_all, "loss/d/total": loss_disc, "learning_rate": optim_g.param_groups[0]["lr"], "grad/norm_d": grad_norm_d, "grad/norm_g": grad_norm_g, "loss/g/fm": loss_fm, "loss/g/mel": loss_mel, "loss/g/kl": loss_kl}
-                scalar_dict.update({f"loss/g/{i}": v for i, v in enumerate(losses_gen)})
-                scalar_dict.update({f"loss/d_r/{i}": v for i, v in enumerate(losses_disc_r)})
-                scalar_dict.update({f"loss/d_g/{i}": v for i, v in enumerate(losses_disc_g)})
-
-                with torch.no_grad():
-                    o, *_ = net_g.module.infer(*reference) if hasattr(net_g, "module") else net_g.infer(*reference)
-
-                summarize(writer=writer, global_step=global_step, images={"slice/mel_org": plot_spectrogram_to_numpy(y_mel[0].data.cpu().numpy()), "slice/mel_gen": plot_spectrogram_to_numpy(y_hat_mel[0].data.cpu().numpy()), "all/mel": plot_spectrogram_to_numpy(mel[0].data.cpu().numpy())}, scalars=scalar_dict, audios={f"gen/audio_{global_step:07d}": o[0, :, :]}, audio_sample_rate=config.data.sample_rate)
-
-            global_step += 1
-            pbar.update(1)
-
-    def check_overtraining(smoothed_loss_history, threshold, epsilon=0.004):
-        if len(smoothed_loss_history) < threshold + 1: return False
-        for i in range(-threshold, -1):
-            if smoothed_loss_history[i + 1] > smoothed_loss_history[i]: return True
-            if abs(smoothed_loss_history[i + 1] - smoothed_loss_history[i]) >= epsilon: return False
-        return True
-
-    def update_exponential_moving_average(smoothed_loss_history, new_value, smoothing=0.987):
-        smoothed_value = new_value if not smoothed_loss_history else (smoothing * smoothed_loss_history[-1] + (1 - smoothing) * new_value)      
-        smoothed_loss_history.append(smoothed_value)
-        return smoothed_value
-
-    def save_to_json(file_path, loss_disc_history, smoothed_loss_disc_history, loss_gen_history, smoothed_loss_gen_history):
-        with open(file_path, "w") as f:
-            json.dump({"loss_disc_history": loss_disc_history, "smoothed_loss_disc_history": smoothed_loss_disc_history, "loss_gen_history": loss_gen_history, "smoothed_loss_gen_history": smoothed_loss_gen_history}, f)
-    
-    model_add, model_del = [], []
-    done = False
-    
-    if rank == 0:
-        if epoch % save_every_epoch == False:
-            checkpoint_suffix = f"{'latest' if save_only_latest else global_step}.pth"
-            save_checkpoint(net_g, optim_g, config.train.learning_rate, epoch, os.path.join(experiment_dir, "G_" + checkpoint_suffix))
-            save_checkpoint(net_d, optim_d, config.train.learning_rate, epoch, os.path.join(experiment_dir, "D_" + checkpoint_suffix))
-            if custom_save_every_weights: model_add.append(os.path.join("assets", "weights", f"{model_name}_{epoch}e_{global_step}s.pth"))
-
-        if overtraining_detector and epoch > 1:
-            current_loss_disc = float(loss_disc)
-            loss_disc_history.append(current_loss_disc)
-            smoothed_value_disc = update_exponential_moving_average(smoothed_loss_disc_history, current_loss_disc)
-            is_overtraining_disc = check_overtraining(smoothed_loss_disc_history, overtraining_threshold * 2)
-
-            if is_overtraining_disc: consecutive_increases_disc += 1
-            else: consecutive_increases_disc = 0
-
-            current_loss_gen = float(lowest_value["value"])
-            loss_gen_history.append(current_loss_gen)
-            smoothed_value_gen = update_exponential_moving_average(smoothed_loss_gen_history, current_loss_gen)
-            is_overtraining_gen = check_overtraining(smoothed_loss_gen_history, overtraining_threshold, 0.01)
-
-            if is_overtraining_gen: consecutive_increases_gen += 1
-            else: consecutive_increases_gen = 0
-
-            if epoch % save_every_epoch == 0: save_to_json(training_file_path, loss_disc_history, smoothed_loss_disc_history, loss_gen_history, smoothed_loss_gen_history)
-
-            if (is_overtraining_gen and consecutive_increases_gen == overtraining_threshold or is_overtraining_disc and consecutive_increases_disc == (overtraining_threshold * 2)):
-                logger.info(translations["overtraining_find"].format(epoch=epoch, smoothed_value_gen=f"{smoothed_value_gen:.3f}", smoothed_value_disc=f"{smoothed_value_disc:.3f}"))
-                done = True
-            else:
-                logger.info(translations["best_epoch"].format(epoch=epoch, smoothed_value_gen=f"{smoothed_value_gen:.3f}", smoothed_value_disc=f"{smoothed_value_disc:.3f}"))
-                for file in glob.glob(os.path.join("assets", "weights", f"{model_name}_*e_*s_best_epoch.pth")):
-                    model_del.append(file)
-
-                model_add.append(os.path.join("assets", "weights", f"{model_name}_{epoch}e_{global_step}s_best_epoch.pth"))
-        
-        if epoch >= custom_total_epoch:
-            logger.info(translations["success_training"].format(epoch=epoch, global_step=global_step, loss_gen_all=round(loss_gen_all.item(), 3)))
-            logger.info(translations["training_info"].format(lowest_value_rounded=round(float(lowest_value["value"]), 3), lowest_value_epoch=lowest_value['epoch'], lowest_value_step=lowest_value['step']))
-
-            pid_file_path = os.path.join(experiment_dir, "config.json")
-            with open(pid_file_path, "r") as pid_file:
-                pid_data = json.load(pid_file)
-
-            with open(pid_file_path, "w") as pid_file:
-                pid_data.pop("process_pids", None)
-                json.dump(pid_data, pid_file, indent=4)
-
-            if os.path.exists(os.path.join(experiment_dir, "train_pid.txt")): os.remove(os.path.join(experiment_dir, "train_pid.txt"))
-            model_add.append(os.path.join("assets", "weights", f"{model_name}_{epoch}e_{global_step}s.pth"))
-            done = True
-            
-        for m in model_del:
-            os.remove(m)
-        
-        if model_add:
-            ckpt = (net_g.module.state_dict() if hasattr(net_g, "module") else net_g.state_dict())
-            for m in model_add:
-                extract_model(ckpt=ckpt, sr=sample_rate, pitch_guidance=pitch_guidance == True, name=model_name, model_path=m, epoch=epoch, step=global_step, version=version, hps=hps, model_author=model_author, vocoder=vocoder)
-
-        lowest_value_rounded = round(float(lowest_value["value"]), 3)
-
-        if epoch > 1 and overtraining_detector: logger.info(translations["model_training_info"].format(model_name=model_name, epoch=epoch, global_step=global_step, epoch_recorder=epoch_recorder.record(), lowest_value_rounded=lowest_value_rounded, lowest_value_epoch=lowest_value['epoch'], lowest_value_step=lowest_value['step'], remaining_epochs_gen=(overtraining_threshold - consecutive_increases_gen), remaining_epochs_disc=((overtraining_threshold * 2) - consecutive_increases_disc), smoothed_value_gen=f"{smoothed_value_gen:.3f}", smoothed_value_disc=f"{smoothed_value_disc:.3f}"))
-        elif epoch > 1 and overtraining_detector == False: logger.info(translations["model_training_info_2"].format(model_name=model_name, epoch=epoch, global_step=global_step, epoch_recorder=epoch_recorder.record(), lowest_value_rounded=lowest_value_rounded, lowest_value_epoch=lowest_value['epoch'], lowest_value_step=lowest_value['step']))
-        else: logger.info(translations["model_training_info_3"].format(model_name=model_name, epoch=epoch, global_step=global_step, epoch_recorder=epoch_recorder.record()))
-
-        last_loss_gen_all = loss_gen_all
-        if done: os._exit(0)
-
-if __name__ == "__main__": 
-    torch.multiprocessing.set_start_method("spawn")
-    main()

main/library/algorithm/commons.py

@@ -1,60 +0,0 @@
-import torch
-
-
-
-def init_weights(m, mean=0.0, std=0.01):
-    if m.__class__.__name__.find("Conv") != -1: m.weight.data.normal_(mean, std)
-
-def get_padding(kernel_size, dilation=1):
-    return int((kernel_size * dilation - dilation) / 2)
-
-def convert_pad_shape(pad_shape):
-    return [item for sublist in pad_shape[::-1] for item in sublist]
-
-def slice_segments(x, ids_str, segment_size = 4, dim = 2):
-    if dim == 2: ret = torch.zeros_like(x[:, :segment_size])
-    elif dim == 3: ret = torch.zeros_like(x[:, :, :segment_size])
-
-    for i in range(x.size(0)):
-        idx_str = ids_str[i].item()
-        idx_end = idx_str + segment_size
-
-        if dim == 2: ret[i] = x[i, idx_str:idx_end]
-        else: ret[i] = x[i, :, idx_str:idx_end]
-
-    return ret
-
-def rand_slice_segments(x, x_lengths=None, segment_size=4):
-    b, _, t = x.size()
-    if x_lengths is None: x_lengths = t
-
-    ids_str = (torch.rand([b]).to(device=x.device) * (x_lengths - segment_size + 1)).to(dtype=torch.long)
-
-    return slice_segments(x, ids_str, segment_size, dim=3), ids_str
-
-@torch.jit.script
-def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
-    n_channels_int = n_channels[0]
-    
-    in_act = input_a + input_b
-
-    return torch.tanh(in_act[:, :n_channels_int, :]) * torch.sigmoid(in_act[:, n_channels_int:, :])
-
-def sequence_mask(length, max_length = None):
-    if max_length is None: max_length = length.max()
-
-    return torch.arange(max_length, dtype=length.dtype, device=length.device).unsqueeze(0) < length.unsqueeze(1)
-
-def clip_grad_value(parameters, clip_value, norm_type=2):
-    if isinstance(parameters, torch.Tensor): parameters = [parameters]
-    norm_type = float(norm_type)
-
-    if clip_value is not None: clip_value = float(clip_value)
-    total_norm = 0
-
-    for p in list(filter(lambda p: p.grad is not None, parameters)):
-        total_norm += (p.grad.data.norm(norm_type)).item() ** norm_type
-
-        if clip_value is not None: p.grad.data.clamp_(min=-clip_value, max=clip_value)
-
-    return total_norm ** (1.0 / norm_type)

main/library/algorithm/modules.py

@@ -1,60 +0,0 @@
-import os
-import sys
-import torch
-
-sys.path.append(os.getcwd())
-
-from .commons import fused_add_tanh_sigmoid_multiply
-
-class WaveNet(torch.nn.Module):
-    def __init__(self, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0, p_dropout=0):
-        super(WaveNet, self).__init__()
-        assert kernel_size % 2 == 1
-        self.hidden_channels = hidden_channels
-        self.kernel_size = (kernel_size,)
-        self.dilation_rate = dilation_rate
-        self.n_layers = n_layers
-        self.gin_channels = gin_channels
-        self.p_dropout = p_dropout
-        self.in_layers = torch.nn.ModuleList()
-        self.res_skip_layers = torch.nn.ModuleList()
-        self.drop = torch.nn.Dropout(p_dropout)
-        if gin_channels != 0: self.cond_layer = torch.nn.utils.parametrizations.weight_norm(torch.nn.Conv1d(gin_channels, 2 * hidden_channels * n_layers, 1), name="weight")
-        dilations = [dilation_rate ** i for i in range(n_layers)]
-        paddings = [(kernel_size * d - d) // 2 for d in dilations]
-
-        for i in range(n_layers):
-            in_layer = torch.nn.Conv1d(hidden_channels, 2 * hidden_channels, kernel_size, dilation=dilations[i], padding=paddings[i])
-            in_layer = torch.nn.utils.parametrizations.weight_norm(in_layer, name="weight")
-            self.in_layers.append(in_layer)
-            res_skip_channels = (hidden_channels if i == n_layers - 1 else 2 * hidden_channels)
-            res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
-            res_skip_layer = torch.nn.utils.parametrizations.weight_norm(res_skip_layer, name="weight")
-            self.res_skip_layers.append(res_skip_layer)
-
-    def forward(self, x, x_mask, g=None):
-        output = x.clone().zero_()
-        n_channels_tensor = torch.IntTensor([self.hidden_channels])
-
-        if g is not None: g = self.cond_layer(g)
-
-        for i in range(self.n_layers):
-            x_in = self.in_layers[i](x)
-            g_l = (g[:, i * 2 * self.hidden_channels : (i + 1) * 2 * self.hidden_channels, :] if g is not None else 0)
-            res_skip_acts = self.res_skip_layers[i](self.drop(fused_add_tanh_sigmoid_multiply(x_in, g_l, n_channels_tensor)))
-
-            if i < self.n_layers - 1:
-                x = (x + (res_skip_acts[:, : self.hidden_channels, :])) * x_mask
-                output = output + res_skip_acts[:, self.hidden_channels :, :]
-            else: output = output + res_skip_acts
-
-        return output * x_mask
-
-    def remove_weight_norm(self):
-        if self.gin_channels != 0: torch.nn.utils.remove_weight_norm(self.cond_layer)
-
-        for l in self.in_layers:
-            torch.nn.utils.remove_weight_norm(l)
-
-        for l in self.res_skip_layers:
-            torch.nn.utils.remove_weight_norm(l)

main/library/algorithm/mrf_hifigan.py

@@ -1,150 +0,0 @@
-import math
-import torch
-
-import numpy as np
-import torch.nn as nn
-import torch.nn.functional as F
-
-from torch.nn.utils import remove_weight_norm
-from torch.utils.checkpoint import checkpoint
-from torch.nn.utils.parametrizations import weight_norm
-
-LRELU_SLOPE = 0.1
-
-class MRFLayer(nn.Module):
-    def __init__(self, channels, kernel_size, dilation):
-        super().__init__()
-        self.conv1 = weight_norm(nn.Conv1d(channels, channels, kernel_size, padding=(kernel_size * dilation - dilation) // 2, dilation=dilation))
-        self.conv2 = weight_norm(nn.Conv1d(channels, channels, kernel_size, padding=kernel_size // 2, dilation=1))
-
-    def forward(self, x):
-        return x + self.conv2(F.leaky_relu(self.conv1(F.leaky_relu(x, LRELU_SLOPE)), LRELU_SLOPE))
-
-    def remove_weight_norm(self):
-        remove_weight_norm(self.conv1)
-        remove_weight_norm(self.conv2)
-
-class MRFBlock(nn.Module):
-    def __init__(self, channels, kernel_size, dilations):
-        super().__init__()
-        self.layers = nn.ModuleList()
-
-        for dilation in dilations:
-            self.layers.append(MRFLayer(channels, kernel_size, dilation))
-
-    def forward(self, x):
-        for layer in self.layers:
-            x = layer(x)
-
-        return x
-
-    def remove_weight_norm(self):
-        for layer in self.layers:
-            layer.remove_weight_norm()
-
-class SineGenerator(nn.Module):
-    def __init__(self, samp_rate, harmonic_num = 0, sine_amp = 0.1, noise_std = 0.003, voiced_threshold = 0):
-        super(SineGenerator, self).__init__()
-        self.sine_amp = sine_amp
-        self.noise_std = noise_std
-        self.harmonic_num = harmonic_num
-        self.dim = self.harmonic_num + 1
-        self.sampling_rate = samp_rate
-        self.voiced_threshold = voiced_threshold
-
-    def _f02uv(self, f0):
-        return torch.ones_like(f0) * (f0 > self.voiced_threshold)
-
-    def _f02sine(self, f0_values):
-        rad_values = (f0_values / self.sampling_rate) % 1
-        rand_ini = torch.rand(f0_values.shape[0], f0_values.shape[2], dtype=f0_values.dtype, device=f0_values.device)
-        rand_ini[:, 0] = 0
-        rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
-        tmp_over_one = torch.cumsum(rad_values, 1) % 1
-        tmp_over_one_idx = (tmp_over_one[:, 1:, :] - tmp_over_one[:, :-1, :]) < 0
-        cumsum_shift = torch.zeros_like(rad_values)
-        cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
-
-        return torch.sin(torch.cumsum(rad_values + cumsum_shift, dim=1) * 2 * np.pi)
-
-    def forward(self, f0):
-        with torch.no_grad():
-            f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim, dtype=f0.dtype, device=f0.device)
-            f0_buf[:, :, 0] = f0[:, :, 0]
-
-            for idx in np.arange(self.harmonic_num):
-                f0_buf[:, :, idx + 1] = f0_buf[:, :, 0] * (idx + 2)
-
-            sine_waves = self._f02sine(f0_buf) * self.sine_amp
-            uv = self._f02uv(f0)
-            sine_waves = sine_waves * uv + ((uv * self.noise_std + (1 - uv) * self.sine_amp / 3) * torch.randn_like(sine_waves))
-
-        return sine_waves
-
-class SourceModuleHnNSF(nn.Module):
-    def __init__(self, sampling_rate, harmonic_num = 0, sine_amp = 0.1, add_noise_std = 0.003, voiced_threshold = 0):
-        super(SourceModuleHnNSF, self).__init__()
-        self.sine_amp = sine_amp
-        self.noise_std = add_noise_std
-        self.l_sin_gen = SineGenerator(sampling_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshold)
-        self.l_linear = nn.Linear(harmonic_num + 1, 1)
-        self.l_tanh = nn.Tanh()
-
-    def forward(self, x):
-        return self.l_tanh(self.l_linear(self.l_sin_gen(x).to(dtype=self.l_linear.weight.dtype)))
-
-class HiFiGANMRFGenerator(nn.Module):
-    def __init__(self, in_channel, upsample_initial_channel, upsample_rates, upsample_kernel_sizes, resblock_kernel_sizes, resblock_dilations, gin_channels, sample_rate, harmonic_num, checkpointing = False):
-        super().__init__()
-        self.num_kernels = len(resblock_kernel_sizes)
-        self.checkpointing = checkpointing
-        self.f0_upsample = nn.Upsample(scale_factor=np.prod(upsample_rates))
-        self.m_source = SourceModuleHnNSF(sample_rate, harmonic_num)
-        self.conv_pre = weight_norm(nn.Conv1d(in_channel, upsample_initial_channel, kernel_size=7, stride=1, padding=3))
-        self.upsamples = nn.ModuleList()
-        self.noise_convs = nn.ModuleList()
-        stride_f0s = [math.prod(upsample_rates[i + 1 :]) if i + 1 < len(upsample_rates) else 1 for i in range(len(upsample_rates))]
-
-        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
-            self.upsamples.append(weight_norm(nn.ConvTranspose1d(upsample_initial_channel // (2**i), upsample_initial_channel // (2 ** (i + 1)), kernel_size=k, stride=u, padding=((k - u) // 2) if u % 2 == 0 else (u // 2 + u % 2), output_padding=u % 2)))
-            stride = stride_f0s[i]
-            kernel = 1 if stride == 1 else stride * 2 - stride % 2
-            self.noise_convs.append(nn.Conv1d(1, upsample_initial_channel // (2 ** (i + 1)), kernel_size=kernel, stride=stride, padding=0 if stride == 1 else (kernel - stride) // 2))
-
-        self.mrfs = nn.ModuleList()
-        for i in range(len(self.upsamples)):
-            channel = upsample_initial_channel // (2 ** (i + 1))
-            self.mrfs.append(nn.ModuleList([MRFBlock(channel, kernel_size=k, dilations=d) for k, d in zip(resblock_kernel_sizes, resblock_dilations)]))
-
-        self.conv_post = weight_norm(nn.Conv1d(channel, 1, kernel_size=7, stride=1, padding=3))
-        if gin_channels != 0: self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
-
-    def forward(self, x, f0, g = None):
-        har_source = self.m_source(self.f0_upsample(f0[:, None, :]).transpose(-1, -2)).transpose(-1, -2)
-        x = self.conv_pre(x)
-        if g is not None: x += self.cond(g)
-
-        for ups, mrf, noise_conv in zip(self.upsamples, self.mrfs, self.noise_convs):
-            x = F.leaky_relu(x, LRELU_SLOPE)
-
-            if self.training and self.checkpointing:
-                x = checkpoint(ups, x, use_reentrant=False) + noise_conv(har_source)
-                xs = sum([checkpoint(layer, x, use_reentrant=False) for layer in mrf])
-            else:
-                x = ups(x) + noise_conv(har_source)
-                xs = sum([layer(x) for layer in mrf])
-
-            x = xs / self.num_kernels
-
-        return torch.tanh(self.conv_post(F.leaky_relu(x)))
-
-    def remove_weight_norm(self):
-        remove_weight_norm(self.conv_pre)
-
-        for up in self.upsamples:
-            remove_weight_norm(up)
-
-        for mrf in self.mrfs:
-            mrf.remove_weight_norm()
-
-        remove_weight_norm(self.conv_post)

main/library/algorithm/onnx_export.py

@@ -1,50 +0,0 @@
-import os
-import io
-import sys
-import onnx
-import json
-import torch
-import onnxsim
-import warnings
-
-sys.path.append(os.getcwd())
-
-from main.library.algorithm.synthesizers import SynthesizerONNX
-
-warnings.filterwarnings("ignore")
-
-def onnx_exporter(input_path, output_path, is_half=False, device="cpu"):
-    cpt = (torch.load(input_path, map_location="cpu") if os.path.isfile(input_path) else None)
-    cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
-
-    model_name, model_author, epochs, steps, version, f0, model_hash, vocoder, creation_date = cpt.get("model_name", None), cpt.get("author", None), cpt.get("epoch", None), cpt.get("step", None), cpt.get("version", "v1"), cpt.get("f0", 1), cpt.get("model_hash", None), cpt.get("vocoder", "Default"), cpt.get("creation_date", None)
-    text_enc_hidden_dim = 768 if version == "v2" else 256
-    tgt_sr = cpt["config"][-1]
-
-    net_g = SynthesizerONNX(*cpt["config"], use_f0=f0, text_enc_hidden_dim=text_enc_hidden_dim, vocoder=vocoder, checkpointing=False)
-    net_g.load_state_dict(cpt["weight"], strict=False)
-    net_g.eval().to(device)
-    net_g = (net_g.half() if is_half else net_g.float())
-
-    phone = torch.rand(1, 200, text_enc_hidden_dim).to(device)
-    phone_length = torch.tensor([200]).long().to(device)
-    ds = torch.LongTensor([0]).to(device)
-    rnd = torch.rand(1, 192, 200).to(device)
-
-    if f0:
-        args = (phone, phone_length, ds, rnd, torch.randint(size=(1, 200), low=5, high=255).to(device), torch.rand(1, 200).to(device))
-        input_names = ["phone", "phone_lengths", "ds", "rnd", "pitch", "pitchf"]
-        dynamic_axes = {"phone": [1], "rnd": [2], "pitch": [1], "pitchf": [1]}
-    else:
-        args = (phone, phone_length, ds, rnd)
-        input_names = ["phone", "phone_lengths", "ds", "rnd"]
-        dynamic_axes = {"phone": [1], "rnd": [2]}
-
-    with io.BytesIO() as model:
-        torch.onnx.export(net_g, args, model, do_constant_folding=True, opset_version=17, verbose=False, input_names=input_names, output_names=["audio"], dynamic_axes=dynamic_axes)
-
-        model, _ = onnxsim.simplify(onnx.load_model_from_string(model.getvalue()))
-        model.metadata_props.append(onnx.StringStringEntryProto(key="model_info", value=json.dumps({"model_name": model_name, "author": model_author, "epoch": epochs, "step": steps, "version": version, "sr": tgt_sr, "f0": f0, "model_hash": model_hash, "creation_date": creation_date, "vocoder": vocoder, "text_enc_hidden_dim": text_enc_hidden_dim})))
-
-    onnx.save(model, output_path)
-    return output_path

main/library/algorithm/refinegan.py

@@ -1,170 +0,0 @@
-import os
-import sys
-import math
-import torch
-
-import numpy as np
-import torch.nn as nn
-import torch.nn.functional as F
-
-from torch.utils.checkpoint import checkpoint
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
-
-sys.path.append(os.getcwd())
-
-from main.library.algorithm.commons import init_weights, get_padding
-
-
-class ResBlock(nn.Module):
-    def __init__(self, channels, kernel_size = 7, dilation = (1, 3, 5), leaky_relu_slope = 0.2):
-        super().__init__()
-        self.leaky_relu_slope = leaky_relu_slope
-        self.convs1 = nn.ModuleList([weight_norm(nn.Conv1d(channels, channels, kernel_size, stride=1, dilation=d, padding=get_padding(kernel_size, d))) for d in dilation])
-        self.convs1.apply(init_weights)
-        self.convs2 = nn.ModuleList([weight_norm(nn.Conv1d(channels, channels, kernel_size, stride=1, dilation=1, padding=get_padding(kernel_size, 1))) for _ in dilation])
-        self.convs2.apply(init_weights)
-
-    def forward(self, x):
-        for c1, c2 in zip(self.convs1, self.convs2):
-            x = c2(F.leaky_relu(c1(F.leaky_relu(x, self.leaky_relu_slope)), self.leaky_relu_slope)) + x
-
-        return x
-
-    def remove_weight_norm(self):
-        for c1, c2 in zip(self.convs1, self.convs2):
-            remove_weight_norm(c1)
-            remove_weight_norm(c2)
-
-class AdaIN(nn.Module):
-    def __init__(self, *, channels, leaky_relu_slope = 0.2):
-        super().__init__()
-        self.weight = nn.Parameter(torch.ones(channels))
-        self.activation = nn.LeakyReLU(leaky_relu_slope)
-
-    def forward(self, x):
-        return self.activation(x + (torch.randn_like(x) * self.weight[None, :, None]))
-    
-class ParallelResBlock(nn.Module):
-    def __init__(self, *, in_channels, out_channels, kernel_sizes = (3, 7, 11), dilation = (1, 3, 5), leaky_relu_slope = 0.2):
-        super().__init__()
-        self.in_channels = in_channels
-        self.out_channels = out_channels
-        self.input_conv = nn.Conv1d(in_channels=in_channels, out_channels=out_channels, kernel_size=7, stride=1, padding=3)
-        self.input_conv.apply(init_weights)
-        self.blocks = nn.ModuleList([nn.Sequential(AdaIN(channels=out_channels), ResBlock(out_channels, kernel_size=kernel_size, dilation=dilation, leaky_relu_slope=leaky_relu_slope), AdaIN(channels=out_channels)) for kernel_size in kernel_sizes])
-
-    def forward(self, x):
-        x = self.input_conv(x)
-        return torch.stack([block(x) for block in self.blocks], dim=0).mean(dim=0)
-
-    def remove_weight_norm(self):
-        remove_weight_norm(self.input_conv)
-        for block in self.blocks:
-            block[1].remove_weight_norm()
-
-class SineGenerator(nn.Module):
-    def __init__(self, samp_rate, harmonic_num=0, sine_amp=0.1, noise_std=0.003, voiced_threshold=0):
-        super(SineGenerator, self).__init__()
-        self.sine_amp = sine_amp
-        self.noise_std = noise_std
-        self.harmonic_num = harmonic_num
-        self.dim = self.harmonic_num + 1
-        self.sampling_rate = samp_rate
-        self.voiced_threshold = voiced_threshold
-        self.merge = nn.Sequential(nn.Linear(self.dim, 1, bias=False), nn.Tanh())
-
-    def _f02uv(self, f0):
-        return torch.ones_like(f0) * (f0 > self.voiced_threshold)
-    
-    def _f02sine(self, f0_values):
-        rad_values = (f0_values / self.sampling_rate) % 1
-        rand_ini = torch.rand(f0_values.shape[0], f0_values.shape[2], dtype=f0_values.dtype, device=f0_values.device)
-
-        rand_ini[:, 0] = 0
-        rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
-
-        tmp_over_one = torch.cumsum(rad_values, 1) % 1
-        tmp_over_one_idx = (tmp_over_one[:, 1:, :] - tmp_over_one[:, :-1, :]) < 0
-
-        cumsum_shift = torch.zeros_like(rad_values)
-        cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
-
-        return torch.sin(torch.cumsum(rad_values + cumsum_shift, dim=1) * 2 * np.pi)
-    
-    def forward(self, f0):
-        with torch.no_grad():
-            f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim, dtype=f0.dtype, device=f0.device)
-            f0_buf[:, :, 0] = f0[:, :, 0]
-
-            for idx in np.arange(self.harmonic_num):
-                f0_buf[:, :, idx + 1] = f0_buf[:, :, 0] * (idx + 2)
-
-            sine_waves = self._f02sine(f0_buf) * self.sine_amp
-            uv = self._f02uv(f0)
-            sine_waves = sine_waves * uv + ((uv * self.noise_std + (1 - uv) * self.sine_amp / 3) * torch.randn_like(sine_waves))
-
-        return self.merge(sine_waves)
-    
-class RefineGANGenerator(nn.Module):
-    def __init__(self, *, sample_rate = 44100, upsample_rates = (8, 8, 2, 2), leaky_relu_slope = 0.2, num_mels = 128, gin_channels = 256, checkpointing = False, upsample_initial_channel = 512):
-        super().__init__()
-        self.upsample_rates = upsample_rates
-        self.checkpointing = checkpointing
-        self.leaky_relu_slope = leaky_relu_slope
-        self.upp = np.prod(upsample_rates)
-        self.m_source = SineGenerator(sample_rate)
-        self.pre_conv = weight_norm(nn.Conv1d(1, upsample_initial_channel // 2, 7, 1, padding=3))
-        stride_f0s = [math.prod(upsample_rates[i + 1 :]) if i + 1 < len(upsample_rates) else 1 for i in range(len(upsample_rates))]
-
-        channels = upsample_initial_channel
-        self.downsample_blocks = nn.ModuleList([])
-
-        for i, _ in enumerate(upsample_rates):
-            stride = stride_f0s[i]
-            kernel = 1 if stride == 1 else stride * 2 - stride % 2
-
-            self.downsample_blocks.append(weight_norm(nn.Conv1d(1, channels // 2 ** (i + 2), kernel, stride, padding=0 if stride == 1 else (kernel - stride) // 2)))
-
-        self.mel_conv = weight_norm(nn.Conv1d(num_mels, channels // 2, 7, 1, padding=3))
-        self.mel_conv.apply(init_weights)
-
-        if gin_channels != 0: self.cond = nn.Conv1d(256, channels // 2, 1)
-
-        self.upsample_blocks = nn.ModuleList([])
-        self.upsample_conv_blocks = nn.ModuleList([])
-
-        for rate in upsample_rates:
-            new_channels = channels // 2
-            self.upsample_blocks.append(nn.Upsample(scale_factor=rate, mode="linear"))
-            self.upsample_conv_blocks.append(ParallelResBlock(in_channels=channels + channels // 4, out_channels=new_channels, kernel_sizes=(3, 7, 11), dilation=(1, 3, 5), leaky_relu_slope=leaky_relu_slope))
-            channels = new_channels
-
-        self.conv_post = weight_norm(nn.Conv1d(channels, 1, 7, 1, padding=3, bias=False))
-        self.conv_post.apply(init_weights)
-
-    def forward(self, mel, f0, g = None):
-        har_source = self.m_source(F.interpolate(f0.unsqueeze(1), size=mel.shape[-1] * self.upp, mode="linear").transpose(1, 2)).transpose(1, 2)
-        x = F.interpolate(self.pre_conv(har_source), size=mel.shape[-1], mode="linear")
-
-        mel = self.mel_conv(mel)
-        if g is not None: mel += self.cond(g)
-
-        x = torch.cat([mel, x], dim=1)
-
-        for ups, res, down in zip(self.upsample_blocks, self.upsample_conv_blocks, self.downsample_blocks):
-            x = F.leaky_relu(x, self.leaky_relu_slope)
-            x = checkpoint(res, torch.cat([checkpoint(ups, x, use_reentrant=False), down(har_source)], dim=1), use_reentrant=False) if self.training and self.checkpointing else res(torch.cat([ups(x), down(har_source)], dim=1))
-
-        return torch.tanh(self.conv_post(F.leaky_relu(x, self.leaky_relu_slope)))
-
-    def remove_weight_norm(self):
-        remove_weight_norm(self.pre_conv)
-        remove_weight_norm(self.mel_conv)
-        remove_weight_norm(self.conv_post)
-
-        for block in self.downsample_blocks:
-            block.remove_weight_norm()
-
-        for block in self.upsample_conv_blocks:
-            block.remove_weight_norm()

main/library/algorithm/residuals.py

@@ -1,140 +0,0 @@
-import os
-import sys
-import torch
-
-from torch.nn.utils import remove_weight_norm
-from torch.nn.utils.parametrizations import weight_norm
-
-sys.path.append(os.getcwd())
-
-from .modules import WaveNet
-from .commons import get_padding, init_weights
-
-
-LRELU_SLOPE = 0.1
-
-def create_conv1d_layer(channels, kernel_size, dilation):
-    return weight_norm(torch.nn.Conv1d(channels, channels, kernel_size, 1, dilation=dilation, padding=get_padding(kernel_size, dilation)))
-
-def apply_mask(tensor, mask):
-    return tensor * mask if mask is not None else tensor
-
-class ResBlockBase(torch.nn.Module):
-    def __init__(self, channels, kernel_size, dilations):
-        super(ResBlockBase, self).__init__()
-
-        self.convs1 = torch.nn.ModuleList([create_conv1d_layer(channels, kernel_size, d) for d in dilations])
-        self.convs1.apply(init_weights)
-
-        self.convs2 = torch.nn.ModuleList([create_conv1d_layer(channels, kernel_size, 1) for _ in dilations])
-        self.convs2.apply(init_weights)
-
-    def forward(self, x, x_mask=None):
-        for c1, c2 in zip(self.convs1, self.convs2):
-            x = c2(apply_mask(torch.nn.functional.leaky_relu(c1(apply_mask(torch.nn.functional.leaky_relu(x, LRELU_SLOPE), x_mask)), LRELU_SLOPE), x_mask)) + x
-
-        return apply_mask(x, x_mask)
-
-    def remove_weight_norm(self):
-        for conv in self.convs1 + self.convs2:
-            remove_weight_norm(conv)
-
-class ResBlock(ResBlockBase):
-    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
-        super(ResBlock, self).__init__(channels, kernel_size, dilation)
-
-class Log(torch.nn.Module):
-    def forward(self, x, x_mask, reverse=False, **kwargs):
-        if not reverse:
-            y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
-            return y, torch.sum(-y, [1, 2])
-        else: return torch.exp(x) * x_mask
-
-class Flip(torch.nn.Module):
-    def forward(self, x, *args, reverse=False, **kwargs):
-        x = torch.flip(x, [1])
-
-        if not reverse: return x, torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
-        else: return x
-
-class ElementwiseAffine(torch.nn.Module):
-    def __init__(self, channels):
-        super().__init__()
-        self.channels = channels
-        self.m = torch.nn.Parameter(torch.zeros(channels, 1))
-        self.logs = torch.nn.Parameter(torch.zeros(channels, 1))
-
-    def forward(self, x, x_mask, reverse=False, **kwargs):
-        if not reverse: return ((self.m + torch.exp(self.logs) * x) * x_mask), torch.sum(self.logs * x_mask, [1, 2])
-        else: return (x - self.m) * torch.exp(-self.logs) * x_mask
-
-class ResidualCouplingBlock(torch.nn.Module):
-    def __init__(self, channels, hidden_channels, kernel_size, dilation_rate, n_layers, n_flows=4, gin_channels=0):
-        super(ResidualCouplingBlock, self).__init__()
-        self.channels = channels
-        self.hidden_channels = hidden_channels
-        self.kernel_size = kernel_size
-        self.dilation_rate = dilation_rate
-        self.n_layers = n_layers
-        self.n_flows = n_flows
-        self.gin_channels = gin_channels
-        self.flows = torch.nn.ModuleList()
-
-        for _ in range(n_flows):
-            self.flows.append(ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True))
-            self.flows.append(Flip())
-
-    def forward(self, x, x_mask, g = None, reverse = False):
-        if not reverse:
-            for flow in self.flows:
-                x, _ = flow(x, x_mask, g=g, reverse=reverse)
-        else:
-            for flow in reversed(self.flows):
-                x = flow.forward(x, x_mask, g=g, reverse=reverse)
-
-        return x
-
-    def remove_weight_norm(self):
-        for i in range(self.n_flows):
-            self.flows[i * 2].remove_weight_norm()
-
-    def __prepare_scriptable__(self):
-        for i in range(self.n_flows):
-            for hook in self.flows[i * 2]._forward_pre_hooks.values():
-                if (hook.__module__ == "torch.nn.utils.parametrizations.weight_norm" and hook.__class__.__name__ == "WeightNorm"): torch.nn.utils.remove_weight_norm(self.flows[i * 2])
-
-        return self
-
-class ResidualCouplingLayer(torch.nn.Module):
-    def __init__(self, channels, hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=0, gin_channels=0, mean_only=False):
-        assert channels % 2 == 0, "Channels/2"
-        super().__init__()
-        self.channels = channels
-        self.hidden_channels = hidden_channels
-        self.kernel_size = kernel_size
-        self.dilation_rate = dilation_rate
-        self.n_layers = n_layers
-        self.half_channels = channels // 2
-        self.mean_only = mean_only
-
-        self.pre = torch.nn.Conv1d(self.half_channels, hidden_channels, 1)
-        self.enc = WaveNet(hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=p_dropout, gin_channels=gin_channels)
-        self.post = torch.nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
-
-        self.post.weight.data.zero_()
-        self.post.bias.data.zero_()
-
-    def forward(self, x, x_mask, g=None, reverse=False):
-        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
-        stats = self.post(self.enc((self.pre(x0) * x_mask), x_mask, g=g)) * x_mask
-
-        if not self.mean_only: m, logs = torch.split(stats, [self.half_channels] * 2, 1)
-        else:
-            m = stats
-            logs = torch.zeros_like(m)
-
-        if not reverse: return torch.cat([x0, (m + x1 * torch.exp(logs) * x_mask)], 1), torch.sum(logs, [1, 2])
-        else: return torch.cat([x0, ((x1 - m) * torch.exp(-logs) * x_mask)], 1)
-
-    def remove_weight_norm(self):
-        self.enc.remove_weight_norm()

main/library/algorithm/separator.py

@@ -1,320 +0,0 @@
-import os
-import sys
-import time
-import yaml
-import torch
-import codecs
-import hashlib
-import logging
-import platform
-import warnings
-import requests
-import onnxruntime
-
-from importlib import metadata, import_module
-
-now_dir = os.getcwd()
-sys.path.append(now_dir)
-
-from main.configs.config import Config
-from main.tools.huggingface import HF_download_file
-
-translations = Config().translations
-
-
-class Separator:
-    def __init__(self, logger=logging.getLogger(__name__), log_level=logging.INFO, log_formatter=None, model_file_dir="assets/models/uvr5", output_dir=None, output_format="wav", output_bitrate=None, normalization_threshold=0.9, output_single_stem=None, invert_using_spec=False, sample_rate=44100, mdx_params={"hop_length": 1024, "segment_size": 256, "overlap": 0.25, "batch_size": 1, "enable_denoise": False}, demucs_params={"segment_size": "Default", "shifts": 2, "overlap": 0.25, "segments_enabled": True}):
-        self.logger = logger
-        self.log_level = log_level
-        self.log_formatter = log_formatter
-        self.log_handler = logging.StreamHandler()
-
-        if self.log_formatter is None: self.log_formatter = logging.Formatter("%(asctime)s - %(levelname)s - %(module)s - %(message)s")
-        self.log_handler.setFormatter(self.log_formatter)
-
-        if not self.logger.hasHandlers(): self.logger.addHandler(self.log_handler)
-        if log_level > logging.DEBUG: warnings.filterwarnings("ignore")
-
-        self.logger.info(translations["separator_info"].format(output_dir=output_dir, output_format=output_format))
-        self.model_file_dir = model_file_dir
-
-        if output_dir is None:
-            output_dir = now_dir
-            self.logger.info(translations["output_dir_is_none"])
-
-        self.output_dir = output_dir
-
-        os.makedirs(self.model_file_dir, exist_ok=True)
-        os.makedirs(self.output_dir, exist_ok=True)
-
-        self.output_format = output_format
-        self.output_bitrate = output_bitrate
-
-        if self.output_format is None: self.output_format = "wav"
-        self.normalization_threshold = normalization_threshold
-        if normalization_threshold <= 0 or normalization_threshold > 1: raise ValueError(translations[">0or=1"])
-
-        self.output_single_stem = output_single_stem
-        if output_single_stem is not None: self.logger.debug(translations["output_single"].format(output_single_stem=output_single_stem))
-
-        self.invert_using_spec = invert_using_spec
-        if self.invert_using_spec: self.logger.debug(translations["step2"])
-
-        self.sample_rate = int(sample_rate)
-        self.arch_specific_params = {"MDX": mdx_params, "Demucs": demucs_params}
-        self.torch_device = None
-        self.torch_device_cpu = None
-        self.torch_device_mps = None
-        self.onnx_execution_provider = None
-        self.model_instance = None
-        self.model_is_uvr_vip = False
-        self.model_friendly_name = None
-        self.setup_accelerated_inferencing_device()
-
-    def setup_accelerated_inferencing_device(self):
-        system_info = self.get_system_info()
-        self.log_onnxruntime_packages()
-        self.setup_torch_device(system_info)
-
-    def get_system_info(self):
-        os_name = platform.system()
-        os_version = platform.version()
-        self.logger.info(f"{translations['os']}: {os_name} {os_version}")
-        system_info = platform.uname()
-        self.logger.info(translations["platform_info"].format(system_info=system_info, node=system_info.node, release=system_info.release, machine=system_info.machine, processor=system_info.processor))
-        python_version = platform.python_version()
-        self.logger.info(f"{translations['name_ver'].format(name='python')}: {python_version}")
-        pytorch_version = torch.__version__
-        self.logger.info(f"{translations['name_ver'].format(name='pytorch')}: {pytorch_version}")
-
-        return system_info
-
-    def log_onnxruntime_packages(self):
-        onnxruntime_gpu_package = self.get_package_distribution("onnxruntime-gpu")
-        onnxruntime_cpu_package = self.get_package_distribution("onnxruntime")
-
-        if onnxruntime_gpu_package is not None: self.logger.info(f"{translations['install_onnx'].format(pu='GPU')}: {onnxruntime_gpu_package.version}")
-        if onnxruntime_cpu_package is not None: self.logger.info(f"{translations['install_onnx'].format(pu='CPU')}: {onnxruntime_cpu_package.version}")
-
-    def setup_torch_device(self, system_info):
-        hardware_acceleration_enabled = False
-        ort_providers = onnxruntime.get_available_providers()
-        self.torch_device_cpu = torch.device("cpu")
-
-        if torch.cuda.is_available():
-            self.configure_cuda(ort_providers)
-            hardware_acceleration_enabled = True
-        elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available() and system_info.processor == "arm":
-            self.configure_mps(ort_providers)
-            hardware_acceleration_enabled = True
-
-        if not hardware_acceleration_enabled:
-            self.logger.info(translations["running_in_cpu"])
-            self.torch_device = self.torch_device_cpu
-            self.onnx_execution_provider = ["CPUExecutionProvider"]
-
-    def configure_cuda(self, ort_providers):
-        self.logger.info(translations["running_in_cuda"])
-        self.torch_device = torch.device("cuda")
-
-        if "CUDAExecutionProvider" in ort_providers:
-            self.logger.info(translations["onnx_have"].format(have='CUDAExecutionProvider'))
-            self.onnx_execution_provider = ["CUDAExecutionProvider"]
-        else: self.logger.warning(translations["onnx_not_have"].format(have='CUDAExecutionProvider'))
-
-    def configure_mps(self, ort_providers):
-        self.logger.info(translations["set_torch_mps"])
-        self.torch_device_mps = torch.device("mps")
-        self.torch_device = self.torch_device_mps
-
-        if "CoreMLExecutionProvider" in ort_providers:
-            self.logger.info(translations["onnx_have"].format(have='CoreMLExecutionProvider'))
-            self.onnx_execution_provider = ["CoreMLExecutionProvider"]
-        else: self.logger.warning(translations["onnx_not_have"].format(have='CoreMLExecutionProvider'))
-
-    def get_package_distribution(self, package_name):
-        try:
-            return metadata.distribution(package_name)
-        except metadata.PackageNotFoundError:
-            self.logger.debug(translations["python_not_install"].format(package_name=package_name))
-            return None
-
-    def get_model_hash(self, model_path):
-        self.logger.debug(translations["hash"].format(model_path=model_path))
-
-        try:
-            with open(model_path, "rb") as f:
-                f.seek(-10000 * 1024, 2)
-                return hashlib.md5(f.read()).hexdigest()
-        except IOError as e:
-            self.logger.error(translations["ioerror"].format(e=e))
-            return hashlib.md5(open(model_path, "rb").read()).hexdigest()
-
-    def download_file_if_not_exists(self, url, output_path):
-        if os.path.isfile(output_path):
-            self.logger.debug(translations["cancel_download"].format(output_path=output_path))
-            return
-
-        self.logger.debug(translations["download_model"].format(url=url, output_path=output_path))
-        HF_download_file(url, output_path)
-
-    def print_uvr_vip_message(self):
-        if self.model_is_uvr_vip:
-            self.logger.warning(translations["vip_model"].format(model_friendly_name=self.model_friendly_name))
-            self.logger.warning(translations["vip_print"])
-
-    def list_supported_model_files(self):
-        response = requests.get(codecs.decode("uggcf://uhttvatsnpr.pb/NauC/Ivrganzrfr-EIP-Cebwrpg/enj/znva/wfba/hie_zbqryf.wfba", "rot13"))
-        response.raise_for_status()
-        model_downloads_list = response.json()
-        self.logger.debug(translations["load_download_json"])
-
-        return {"MDX": {**model_downloads_list["mdx_download_list"], **model_downloads_list["mdx_download_vip_list"]}, "Demucs": {key: value for key, value in model_downloads_list["demucs_download_list"].items() if key.startswith("Demucs v4")}}
-    
-    def download_model_files(self, model_filename):
-        model_path = os.path.join(self.model_file_dir, model_filename)
-        supported_model_files_grouped = self.list_supported_model_files()
-
-        yaml_config_filename = None
-        self.logger.debug(translations["search_model"].format(model_filename=model_filename))
-
-        for model_type, model_list in supported_model_files_grouped.items():
-            for model_friendly_name, model_download_list in model_list.items():
-                self.model_is_uvr_vip = "VIP" in model_friendly_name
-                model_repo_url_prefix = codecs.decode("uggcf://uhttvatsnpr.pb/NauC/Ivrganzrfr-EIP-Cebwrpg/erfbyir/znva/hie5_zbqryf", "rot13")
-
-                if isinstance(model_download_list, str) and model_download_list == model_filename:
-                    self.logger.debug(translations["single_model"].format(model_friendly_name=model_friendly_name))
-                    self.model_friendly_name = model_friendly_name
-
-                    try:
-                        self.download_file_if_not_exists(f"{model_repo_url_prefix}/MDX/{model_filename}", model_path)
-                    except RuntimeError:
-                        self.logger.warning(translations["not_found_model"])
-                        self.download_file_if_not_exists(f"{model_repo_url_prefix}/Demucs/{model_filename}", model_path)
-
-                    self.print_uvr_vip_message()
-                    self.logger.debug(translations["single_model_path"].format(model_path=model_path))
-
-                    return model_filename, model_type, model_friendly_name, model_path, yaml_config_filename
-                elif isinstance(model_download_list, dict):
-                    this_model_matches_input_filename = False
-
-                    for file_name, file_url in model_download_list.items():
-                        if file_name == model_filename or file_url == model_filename:
-                            self.logger.debug(translations["find_model"].format(model_filename=model_filename, model_friendly_name=model_friendly_name))
-                            this_model_matches_input_filename = True
-
-                    if this_model_matches_input_filename:
-                        self.logger.debug(translations["find_models"].format(model_friendly_name=model_friendly_name))
-                        self.model_friendly_name = model_friendly_name
-                        self.print_uvr_vip_message()
-
-                        for config_key, config_value in model_download_list.items():
-                            self.logger.debug(f"{translations['find_path']}: {config_key} -> {config_value}")
-
-                            if config_value.startswith("http"): self.download_file_if_not_exists(config_value, os.path.join(self.model_file_dir, config_key))
-                            elif config_key.endswith(".ckpt"):
-                                try:
-                                    self.download_file_if_not_exists(f"{model_repo_url_prefix}/Demucs/{config_key}", os.path.join(self.model_file_dir, config_key))
-                                except RuntimeError:
-                                    self.logger.warning(translations["not_found_model_warehouse"])
-
-                                if model_filename.endswith(".yaml"):
-                                    self.logger.warning(translations["yaml_warning"].format(model_filename=model_filename))
-                                    self.logger.warning(translations["yaml_warning_2"].format(config_key=config_key))
-                                    self.logger.warning(translations["yaml_warning_3"])
-
-                                    model_filename = config_key
-                                    model_path = os.path.join(self.model_file_dir, f"{model_filename}")
-
-                                yaml_config_filename = config_value
-                                yaml_config_filepath = os.path.join(self.model_file_dir, yaml_config_filename)
-
-                                try:
-                                    self.download_file_if_not_exists(f"{model_repo_url_prefix}/mdx_c_configs/{yaml_config_filename}", yaml_config_filepath)
-                                except RuntimeError:
-                                    self.logger.debug(translations["yaml_debug"])
-                            else: self.download_file_if_not_exists(f"{model_repo_url_prefix}/Demucs/{config_value}", os.path.join(self.model_file_dir, config_value))
-
-                        self.logger.debug(translations["download_model_friendly"].format(model_friendly_name=model_friendly_name, model_path=model_path))
-                        return model_filename, model_type, model_friendly_name, model_path, yaml_config_filename
-
-        raise ValueError(translations["not_found_model_2"].format(model_filename=model_filename))
-
-    def load_model_data_from_yaml(self, yaml_config_filename):
-        model_data_yaml_filepath = os.path.join(self.model_file_dir, yaml_config_filename) if not os.path.exists(yaml_config_filename) else yaml_config_filename
-        self.logger.debug(translations["load_yaml"].format(model_data_yaml_filepath=model_data_yaml_filepath))
-        
-        model_data = yaml.load(open(model_data_yaml_filepath, encoding="utf-8"), Loader=yaml.FullLoader)
-        self.logger.debug(translations["load_yaml_2"].format(model_data=model_data))
-
-        if "roformer" in model_data_yaml_filepath: model_data["is_roformer"] = True
-        return model_data
-
-    def load_model_data_using_hash(self, model_path):
-        self.logger.debug(translations["hash_md5"])
-        model_hash = self.get_model_hash(model_path)
-
-        self.logger.debug(translations["model_hash"].format(model_path=model_path, model_hash=model_hash))
-        mdx_model_data_path = codecs.decode("uggcf://uhttvatsnpr.pb/NauC/Ivrganzrfr-EIP-Cebwrpg/enj/znva/wfba/zbqry_qngn.wfba", "rot13")
-        self.logger.debug(translations["mdx_data"].format(mdx_model_data_path=mdx_model_data_path))
-
-        response = requests.get(mdx_model_data_path)
-        response.raise_for_status()
-
-        mdx_model_data_object = response.json()
-        self.logger.debug(translations["load_mdx"])
-
-        if model_hash in mdx_model_data_object: model_data = mdx_model_data_object[model_hash]
-        else: raise ValueError(translations["model_not_support"].format(model_hash=model_hash))
-
-        self.logger.debug(translations["uvr_json"].format(model_hash=model_hash, model_data=model_data))
-        return model_data
-
-    def load_model(self, model_filename):
-        self.logger.info(translations["loading_model"].format(model_filename=model_filename))
-        load_model_start_time = time.perf_counter()
-        model_filename, model_type, model_friendly_name, model_path, yaml_config_filename = self.download_model_files(model_filename)
-        self.logger.debug(translations["download_model_friendly_2"].format(model_friendly_name=model_friendly_name, model_path=model_path))
-
-        if model_path.lower().endswith(".yaml"): yaml_config_filename = model_path
-
-        common_params = {"logger": self.logger, "log_level": self.log_level, "torch_device": self.torch_device, "torch_device_cpu": self.torch_device_cpu, "torch_device_mps": self.torch_device_mps, "onnx_execution_provider": self.onnx_execution_provider, "model_name": model_filename.split(".")[0], "model_path": model_path, "model_data": self.load_model_data_from_yaml(yaml_config_filename) if yaml_config_filename is not None else self.load_model_data_using_hash(model_path), "output_format": self.output_format, "output_bitrate": self.output_bitrate, "output_dir": self.output_dir, "normalization_threshold": self.normalization_threshold, "output_single_stem": self.output_single_stem, "invert_using_spec": self.invert_using_spec, "sample_rate": self.sample_rate}
-        separator_classes = {"MDX": "mdx_separator.MDXSeparator", "Demucs": "demucs_separator.DemucsSeparator"}
-
-        if model_type not in self.arch_specific_params or model_type not in separator_classes: raise ValueError(translations["model_type_not_support"].format(model_type=model_type))
-        if model_type == "Demucs" and sys.version_info < (3, 10): raise Exception(translations["demucs_not_support_python<3.10"])
-
-        self.logger.debug(f"{translations['import_module']} {model_type}: {separator_classes[model_type]}")
-        module_name, class_name = separator_classes[model_type].split(".")
-        separator_class = getattr(import_module(f"main.library.architectures.{module_name}"), class_name)
-
-        self.logger.debug(f"{translations['initialization']} {model_type}: {separator_class}")
-        self.model_instance = separator_class(common_config=common_params, arch_config=self.arch_specific_params[model_type])
-
-        self.logger.debug(translations["loading_model_success"])
-        self.logger.info(f"{translations['loading_model_duration']}: {time.strftime('%H:%M:%S', time.gmtime(int(time.perf_counter() - load_model_start_time)))}")
-
-    def separate(self, audio_file_path):
-        self.logger.info(f"{translations['starting_separator']}: {audio_file_path}")
-        separate_start_time = time.perf_counter()
-
-        self.logger.debug(translations["normalization"].format(normalization_threshold=self.normalization_threshold))
-        output_files = self.model_instance.separate(audio_file_path)
-
-        self.model_instance.clear_gpu_cache()
-        self.model_instance.clear_file_specific_paths()
-
-        self.print_uvr_vip_message()
-
-        self.logger.debug(translations["separator_success_3"])
-        self.logger.info(f"{translations['separator_duration']}: {time.strftime('%H:%M:%S', time.gmtime(int(time.perf_counter() - separate_start_time)))}")
-        return output_files
-
-    def download_model_and_data(self, model_filename):
-        self.logger.info(translations["loading_separator_model"].format(model_filename=model_filename))
-        model_filename, model_type, model_friendly_name, model_path, yaml_config_filename = self.download_model_files(model_filename)
-
-        if model_path.lower().endswith(".yaml"): yaml_config_filename = model_path
-        self.logger.info(translations["downloading_model"].format(model_type=model_type, model_friendly_name=model_friendly_name, model_path=model_path, model_data_dict_size=len(self.load_model_data_from_yaml(yaml_config_filename) if yaml_config_filename is not None else self.load_model_data_using_hash(model_path))))

main/library/algorithm/stftpitchshift.py

@@ -1,250 +0,0 @@
-import numpy as np
-
-from numpy.lib.stride_tricks import sliding_window_view
-
-def istft(frames, framesize, hopsize):
-    frames = np.atleast_2d(frames)
-    assert frames.ndim == 2
-
-    analysis_window_size = np.ravel(framesize)[0]
-    synthesis_window_size = np.ravel(framesize)[-1]
-
-    assert analysis_window_size >= synthesis_window_size
-
-    A = asymmetric_analysis_window(analysis_window_size, synthesis_window_size) if analysis_window_size != synthesis_window_size else symmetric_window(analysis_window_size)
-    S = asymmetric_synthesis_window(analysis_window_size, synthesis_window_size) if analysis_window_size != synthesis_window_size else symmetric_window(synthesis_window_size)
-
-    W = S * hopsize / np.sum(A * S)
-    N = frames.shape[0] * hopsize + analysis_window_size
-
-    y = np.zeros((N), float)
-
-    frames[:,  0] = 0
-    frames[:, -1] = 0
-    frames0 = sliding_window_view(y, analysis_window_size, writeable=True)[::hopsize]
-    frames1 = np.fft.irfft(frames, axis=-1, norm='forward') * W
-
-    for i in range(min(len(frames0), len(frames1))):
-        frames0[i] += frames1[i]
-
-    return y
-
-def asymmetric_synthesis_window(analysis_window_size, synthesis_window_size):
-    n = analysis_window_size
-    m = synthesis_window_size // 2
-
-    right = symmetric_window(2 * m)
-    window = np.zeros(n)
-
-    window[n-m-m:n-m] = np.square(right[:m]) / symmetric_window(2 * n - 2 * m)[n-m-m:n-m]
-    window[-m:] = right[-m:]
-
-    return window
-
-def asymmetric_analysis_window(analysis_window_size, synthesis_window_size):
-    n = analysis_window_size
-    m = synthesis_window_size // 2
-
-    window = np.zeros(n)
-    window[:n-m] = symmetric_window(2 * n - 2 * m)[:n-m]
-    window[-m:] = symmetric_window(2 * m)[-m:]
-
-    return window
-
-def symmetric_window(symmetric_window_size):
-    n = symmetric_window_size
-    window = 0.5 - 0.5 * np.cos(2 * np.pi * np.arange(n) / n)
-
-    return window
-
-def stft(x, framesize, hopsize):
-    x = np.atleast_1d(x)
-    assert x.ndim == 1
-
-    analysis_window_size = np.ravel(framesize)[0]
-    synthesis_window_size = np.ravel(framesize)[-1]
-
-    assert analysis_window_size >= synthesis_window_size
-
-    W = asymmetric_analysis_window(analysis_window_size, synthesis_window_size) if analysis_window_size != synthesis_window_size else symmetric_window(analysis_window_size)
-
-    frames0 = sliding_window_view(x, analysis_window_size, writeable=False)[::hopsize]
-    frames1 = np.fft.rfft(frames0 * W, axis=-1, norm='forward')
-
-    return frames1
-
-def normalize(frames, frames0):
-    for i in range(len(frames)):
-        a = np.real(frames0[i])
-        b = np.real(frames[i])
-        a = np.dot(a, a)
-        b = np.dot(b, b)
-
-        if b == 0: continue
-        frames[i] = np.real(frames[i]) * np.sqrt(a / b) + 1j * np.imag(frames[i])
-
-    return frames
-
-def lowpass(cepstrum, quefrency):
-    cepstrum[1:quefrency] *= 2
-    cepstrum[quefrency+1:] = 0
-
-    return cepstrum
-
-def lifter(frames, quefrency):
-    envelopes = np.zeros(frames.shape)
-
-    for i, frame in enumerate(frames):
-        with np.errstate(divide='ignore', invalid='ignore'):
-            spectrum = np.log10(np.real(frame)) 
-
-        envelopes[i] = np.power(10, np.real(np.fft.rfft(lowpass(np.fft.irfft(spectrum, norm='forward'), quefrency), norm='forward')))
-
-    return envelopes
-
-def resample(x, factor):
-    if factor == 1: return x.copy()
-    y = np.zeros(x.shape, dtype=x.dtype)
-    
-    n = len(x)
-    m = int(n * factor)
-
-    i = np.arange(min(n, m))
-    k = i * (n / m)
-
-    j = np.trunc(k).astype(int)
-    k = k - j
-
-    ok = (0 <= j) & (j < n - 1)
-    y[i[ok]] = k[ok] * x[j[ok] + 1] + (1 - k[ok]) * x[j[ok]]
-
-    return y
-
-def shiftpitch(frames, factors, samplerate):
-    for i in range(len(frames)):
-        magnitudes = np.vstack([resample(np.real(frames[i]), factor) for factor in factors])
-        frequencies = np.vstack([resample(np.imag(frames[i]), factor) * factor for factor in factors])
-
-        magnitudes[(frequencies <= 0) | (frequencies >= samplerate / 2)] = 0
-        mask = np.argmax(magnitudes, axis=0)
-
-        magnitudes = np.take_along_axis(magnitudes, mask[None,:], axis=0)
-        frequencies = np.take_along_axis(frequencies, mask[None,:], axis=0)
-
-        frames[i] = magnitudes + 1j * frequencies
-
-    return frames
-
-def wrap(x):
-    return (x + np.pi) % (2 * np.pi) - np.pi
-
-def encode(frames, framesize, hopsize, samplerate):
-    M, N = frames.shape
-    analysis_framesize = np.ravel(framesize)[0]
-
-    freqinc = samplerate / analysis_framesize
-    phaseinc = 2 * np.pi * hopsize / analysis_framesize
-
-    buffer = np.zeros(N)
-    data = np.zeros((M, N), complex)
-
-    for m, frame in enumerate(frames):
-        arg = np.angle(frame)
-        delta = arg - buffer
-        
-        buffer = arg
-
-        i = np.arange(N)
-        data[m] = np.abs(frame) + 1j * ((i + (wrap(delta - i * phaseinc) / phaseinc)) * freqinc)
-
-    return data
-
-def decode(frames, framesize, hopsize, samplerate):
-    M, N = frames.shape
-    analysis_framesize = np.ravel(framesize)[0]
-    synthesis_framesize = np.ravel(framesize)[-1]
-
-    freqinc = samplerate / analysis_framesize
-    phaseinc = 2 * np.pi * hopsize / analysis_framesize
-    timeshift = 2 * np.pi * synthesis_framesize * np.arange(N) / N if synthesis_framesize != analysis_framesize else 0
-
-    buffer = np.zeros(N)
-    data = np.zeros((M, N), complex)
-
-    for m, frame in enumerate(frames):
-        i = np.arange(N)
-        delta = (i + ((np.imag(frame) - i * freqinc) / freqinc)) * phaseinc
-        buffer += delta
-        arg = buffer.copy()
-        arg -= timeshift 
-        data[m] = np.real(frame) * np.exp(1j * arg)
-
-    return data
-
-class StftPitchShift:
-    def __init__(self, framesize, hopsize, samplerate):
-        self.framesize = framesize
-        self.hopsize = hopsize
-        self.samplerate = samplerate
-
-    def shiftpitch(self, input, factors = 1, quefrency = 0, distortion = 1, normalization = False):
-        input = np.atleast_1d(input)
-        dtype = input.dtype
-        shape = input.shape
-
-        input = np.squeeze(input)
-        if input.ndim != 1: raise ValueError('input.ndim != 1')
-
-        if np.issubdtype(dtype, np.integer):
-            a, b = np.iinfo(dtype).min, np.iinfo(dtype).max
-            input = ((input.astype(float) - a) / (b - a)) * 2 - 1
-        elif not np.issubdtype(dtype, np.floating): raise TypeError('not np.issubdtype(dtype, np.floating)')
-
-        def isnotnormal(x):
-            return (np.isinf(x)) | (np.isnan(x)) | (abs(x) < np.finfo(x.dtype).tiny)
-
-        framesize = self.framesize
-        hopsize = self.hopsize
-        samplerate = self.samplerate
-
-        factors = np.asarray(factors).flatten()
-        quefrency = int(quefrency * samplerate)
-
-        frames = encode(stft(input, framesize, hopsize), framesize, hopsize, samplerate)
-
-        if normalization: frames0 = frames.copy()
-
-        if quefrency:
-            envelopes = lifter(frames, quefrency)
-            mask = isnotnormal(envelopes)
-
-            frames.real /= envelopes
-            frames.real[mask] = 0
-
-            if distortion != 1:
-                envelopes[mask] = 0
-
-                for i in range(len(envelopes)):
-                    envelopes[i] = resample(envelopes[i], distortion)
-
-                mask = isnotnormal(envelopes)
-
-            frames = shiftpitch(frames, factors, samplerate)
-            frames.real *= envelopes
-            frames.real[mask] = 0
-        else: frames = shiftpitch(frames, factors, samplerate)
-
-        if normalization: frames = normalize(frames, frames0)
-
-        output = istft(decode(frames, framesize, hopsize, samplerate), framesize, hopsize)
-        output.resize(shape, refcheck=False)
-
-        if np.issubdtype(dtype, np.integer):
-            a, b = np.iinfo(dtype).min, np.iinfo(dtype).max
-            output = (((output + 1) / 2) * (b - a) + a).clip(a, b).astype(dtype)
-        elif output.dtype != dtype: output = output.astype(dtype)
-
-        assert output.dtype == dtype
-        assert output.shape == shape
-
-        return output

main/library/algorithm/synthesizers.py

@@ -1,490 +0,0 @@
-import os
-import sys
-import math
-import torch
-import numpy as np
-import torch.nn.functional as F
-
-from torch.nn.utils import remove_weight_norm
-from torch.utils.checkpoint import checkpoint
-from torch.nn.utils.parametrizations import weight_norm
-
-sys.path.append(os.getcwd())
-
-from .modules import WaveNet
-from .refinegan import RefineGANGenerator
-from .mrf_hifigan import HiFiGANMRFGenerator
-from .residuals import ResidualCouplingBlock, ResBlock, LRELU_SLOPE
-from .commons import init_weights, slice_segments, rand_slice_segments, sequence_mask, convert_pad_shape
-
-
-class Generator(torch.nn.Module):
-    def __init__(self, initial_channel, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=0):
-        super(Generator, self).__init__()
-        self.num_kernels = len(resblock_kernel_sizes)
-        self.num_upsamples = len(upsample_rates)
-        self.conv_pre = torch.nn.Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
-        self.ups_and_resblocks = torch.nn.ModuleList()
-
-        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
-            self.ups_and_resblocks.append(weight_norm(torch.nn.ConvTranspose1d(upsample_initial_channel // (2**i), upsample_initial_channel // (2 ** (i + 1)), k, u, padding=(k - u) // 2)))
-            ch = upsample_initial_channel // (2 ** (i + 1))
-            for _, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
-                self.ups_and_resblocks.append(ResBlock(ch, k, d))
-
-        self.conv_post = torch.nn.Conv1d(ch, 1, 7, 1, padding=3, bias=False)
-        self.ups_and_resblocks.apply(init_weights)
-        if gin_channels != 0: self.cond = torch.nn.Conv1d(gin_channels, upsample_initial_channel, 1)
-
-        def forward(self, x, g = None):
-            x = self.conv_pre(x)
-            if g is not None: x = x + self.cond(g)
-            
-            resblock_idx = 0
-
-            for _ in range(self.num_upsamples):
-                x = self.ups_and_resblocks[resblock_idx](F.leaky_relu(x, LRELU_SLOPE))
-                resblock_idx += 1
-                xs = 0
-
-                for _ in range(self.num_kernels):
-                    xs += self.ups_and_resblocks[resblock_idx](x)
-                    resblock_idx += 1
-
-                x = xs / self.num_kernels
-
-            return torch.tanh(self.conv_post(F.leaky_relu(x)))
-
-    def __prepare_scriptable__(self):
-        for l in self.ups_and_resblocks:
-            for hook in l._forward_pre_hooks.values():
-                if (hook.__module__ == "torch.nn.utils.parametrizations.weight_norm" and hook.__class__.__name__ == "WeightNorm"): torch.nn.utils.remove_weight_norm(l)
-
-        return self
-    
-    def remove_weight_norm(self):
-        for l in self.ups_and_resblocks:
-            remove_weight_norm(l)
-
-class SineGen(torch.nn.Module):
-    def __init__(self, samp_rate, harmonic_num=0, sine_amp=0.1, noise_std=0.003, voiced_threshold=0, flag_for_pulse=False):
-        super(SineGen, self).__init__()
-        self.sine_amp = sine_amp
-        self.noise_std = noise_std
-        self.harmonic_num = harmonic_num
-        self.dim = self.harmonic_num + 1
-        self.sampling_rate = samp_rate
-        self.voiced_threshold = voiced_threshold
-
-    def _f02uv(self, f0):
-        return torch.ones_like(f0) * (f0 > self.voiced_threshold)
-    
-    def _f02sine(self, f0, upp):
-        rad = f0 / self.sampling_rate * torch.arange(1, upp + 1, dtype=f0.dtype, device=f0.device)
-        rad += F.pad((torch.fmod(rad[:, :-1, -1:].float() + 0.5, 1.0) - 0.5).cumsum(dim=1).fmod(1.0).to(f0), (0, 0, 1, 0), mode='constant')
-        rad = rad.reshape(f0.shape[0], -1, 1)
-        rad *= torch.arange(1, self.dim + 1, dtype=f0.dtype, device=f0.device).reshape(1, 1, -1)
-        rand_ini = torch.rand(1, 1, self.dim, device=f0.device)
-        rand_ini[..., 0] = 0
-        rad += rand_ini
-
-        return torch.sin(2 * np.pi * rad)
-        
-    def forward(self, f0, upp):
-        with torch.no_grad():
-            f0 = f0.unsqueeze(-1)
-            sine_waves = self._f02sine(f0, upp) * self.sine_amp
-            uv = F.interpolate(self._f02uv(f0).transpose(2, 1), scale_factor=float(upp), mode="nearest").transpose(2, 1)
-            sine_waves = sine_waves * uv + ((uv * self.noise_std + (1 - uv) * self.sine_amp / 3) * torch.randn_like(sine_waves))
-
-        return sine_waves
-
-class SourceModuleHnNSF(torch.nn.Module):
-    def __init__(self, sample_rate, harmonic_num=0, sine_amp=0.1, add_noise_std=0.003, voiced_threshod=0):
-        super(SourceModuleHnNSF, self).__init__()
-        self.sine_amp = sine_amp
-        self.noise_std = add_noise_std
-        self.l_sin_gen = SineGen(sample_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshod)
-        self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
-        self.l_tanh = torch.nn.Tanh()
-
-    def forward(self, x, upsample_factor = 1):
-        return self.l_tanh(self.l_linear(self.l_sin_gen(x, upsample_factor).to(dtype=self.l_linear.weight.dtype)))
-
-class GeneratorNSF(torch.nn.Module):
-    def __init__(self, initial_channel, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels, sr, checkpointing = False):
-        super(GeneratorNSF, self).__init__()
-        self.num_kernels = len(resblock_kernel_sizes)
-        self.num_upsamples = len(upsample_rates)
-        self.upp = math.prod(upsample_rates)
-        self.f0_upsamp = torch.nn.Upsample(scale_factor=self.upp)
-        self.m_source = SourceModuleHnNSF(sample_rate=sr, harmonic_num=0)
-
-        self.conv_pre = torch.nn.Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
-        self.checkpointing = checkpointing
-
-        self.ups = torch.nn.ModuleList()
-        self.noise_convs = torch.nn.ModuleList()
-
-        channels = [upsample_initial_channel // (2 ** (i + 1)) for i in range(self.num_upsamples)]
-        stride_f0s = [math.prod(upsample_rates[i + 1 :]) if i + 1 < self.num_upsamples else 1 for i in range(self.num_upsamples)]
-
-        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
-            self.ups.append(weight_norm(torch.nn.ConvTranspose1d(upsample_initial_channel // (2**i), channels[i], k, u, padding=((k - u) // 2) if u % 2 == 0 else (u // 2 + u % 2), output_padding=u % 2)))
-            stride = stride_f0s[i]
-            kernel = 1 if stride == 1 else stride * 2 - stride % 2
-            self.noise_convs.append(torch.nn.Conv1d(1, channels[i], kernel_size=kernel, stride=stride, padding=0 if stride == 1 else (kernel - stride) // 2))
-
-        self.resblocks = torch.nn.ModuleList([ResBlock(channels[i], k, d) for i in range(len(self.ups)) for k, d in zip(resblock_kernel_sizes, resblock_dilation_sizes)])
-        self.conv_post = torch.nn.Conv1d(channels[-1], 1, 7, 1, padding=3, bias=False)
-
-        self.ups.apply(init_weights)
-        if gin_channels != 0: self.cond = torch.nn.Conv1d(gin_channels, upsample_initial_channel, 1)
-
-    def forward(self, x, f0, g = None):
-        har_source = self.m_source(f0, self.upp).transpose(1, 2)
-        x = self.conv_pre(x)
-        if g is not None: x += self.cond(g)
-
-        for i, (ups, noise_convs) in enumerate(zip(self.ups, self.noise_convs)):
-            x = F.leaky_relu(x, LRELU_SLOPE)
-
-            if self.training and self.checkpointing:
-                x = checkpoint(ups, x, use_reentrant=False) + noise_convs(har_source)
-                xs = sum([checkpoint(resblock, x, use_reentrant=False) for j, resblock in enumerate(self.resblocks) if j in range(i * self.num_kernels, (i + 1) * self.num_kernels)])
-            else:
-                x = ups(x) + noise_convs(har_source)
-                xs = sum([resblock(x) for j, resblock in enumerate(self.resblocks) if j in range(i * self.num_kernels, (i + 1) * self.num_kernels)])
-
-            x = xs / self.num_kernels
-
-        return torch.tanh(self.conv_post(F.leaky_relu(x)))
-
-    def remove_weight_norm(self):
-        for l in self.ups:
-            remove_weight_norm(l)
-
-        for l in self.resblocks:
-            l.remove_weight_norm()
-
-class LayerNorm(torch.nn.Module):
-    def __init__(self, channels, eps=1e-5, onnx=False):
-        super().__init__()
-        self.channels = channels
-        self.eps = eps
-        self.onnx = onnx
-        self.gamma = torch.nn.Parameter(torch.ones(channels))
-        self.beta = torch.nn.Parameter(torch.zeros(channels))
-
-    def forward(self, x):
-        x = x.transpose(1, -1)
-        return (F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps) if self.onnx else F.layer_norm(x, (x.size(-1),), self.gamma, self.beta, self.eps)).transpose(1, -1) 
-
-class MultiHeadAttention(torch.nn.Module):
-    def __init__(self, channels, out_channels, n_heads, p_dropout=0.0, window_size=None, heads_share=True, block_length=None, proximal_bias=False, proximal_init=False, onnx=False):
-        super().__init__()
-        assert channels % n_heads == 0
-        self.channels = channels
-        self.out_channels = out_channels
-        self.n_heads = n_heads
-        self.p_dropout = p_dropout
-        self.window_size = window_size
-        self.heads_share = heads_share
-        self.block_length = block_length
-        self.proximal_bias = proximal_bias
-        self.proximal_init = proximal_init
-        self.onnx = onnx
-        self.attn = None
-        self.k_channels = channels // n_heads
-        self.conv_q = torch.nn.Conv1d(channels, channels, 1)
-        self.conv_k = torch.nn.Conv1d(channels, channels, 1)
-        self.conv_v = torch.nn.Conv1d(channels, channels, 1)
-        self.conv_o = torch.nn.Conv1d(channels, out_channels, 1)
-        self.drop = torch.nn.Dropout(p_dropout)
-
-        if window_size is not None:
-            n_heads_rel = 1 if heads_share else n_heads
-            rel_stddev = self.k_channels**-0.5
-
-            self.emb_rel_k = torch.nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
-            self.emb_rel_v = torch.nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
-
-        torch.nn.init.xavier_uniform_(self.conv_q.weight)
-        torch.nn.init.xavier_uniform_(self.conv_k.weight)
-        torch.nn.init.xavier_uniform_(self.conv_v.weight)
-
-        if proximal_init:
-            with torch.no_grad():
-                self.conv_k.weight.copy_(self.conv_q.weight)
-                self.conv_k.bias.copy_(self.conv_q.bias)
-
-    def forward(self, x, c, attn_mask=None):
-        q, k, v = self.conv_q(x), self.conv_k(c), self.conv_v(c)
-        x, self.attn = self.attention(q, k, v, mask=attn_mask)
-
-        return self.conv_o(x)
-
-    def attention(self, query, key, value, mask=None):
-        b, d, t_s, t_t = (*key.size(), query.size(2))
-        query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
-        key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
-        scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
-    
-        if self.window_size is not None:
-            assert (t_s == t_t), "(t_s == t_t)"
-            scores = scores + self._relative_position_to_absolute_position(self._matmul_with_relative_keys(query / math.sqrt(self.k_channels), self._get_relative_embeddings(self.emb_rel_k, t_s, onnx=self.onnx)), onnx=self.onnx)
-
-        if self.proximal_bias:
-            assert t_s == t_t, "t_s == t_t"
-            scores = scores + self._attention_bias_proximal(t_s).to(device=scores.device, dtype=scores.dtype)
-
-        if mask is not None:
-            scores = scores.masked_fill(mask == 0, -1e4)
-            if self.block_length is not None:
-                assert (t_s == t_t), "(t_s == t_t)"
-                scores = scores.masked_fill((torch.ones_like(scores).triu(-self.block_length).tril(self.block_length)) == 0, -1e4)
-
-        p_attn = self.drop(F.softmax(scores, dim=-1))
-        output = torch.matmul(p_attn, value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3))
-
-        if self.window_size is not None: output = output + self._matmul_with_relative_values(self._absolute_position_to_relative_position(p_attn, onnx=self.onnx), self._get_relative_embeddings(self.emb_rel_v, t_s, onnx=self.onnx))
-        return (output.transpose(2, 3).contiguous().view(b, d, t_t)), p_attn
-
-    def _matmul_with_relative_values(self, x, y):
-        return torch.matmul(x, y.unsqueeze(0))
-
-    def _matmul_with_relative_keys(self, x, y):
-        return torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
-
-    def _get_relative_embeddings(self, relative_embeddings, length, onnx=False):
-        if onnx:
-            pad_length = torch.clamp(length - (self.window_size + 1), min=0)
-            slice_start_position = torch.clamp((self.window_size + 1) - length, min=0)
-
-            return (F.pad(relative_embeddings, [0, 0, pad_length, pad_length, 0, 0]) if pad_length > 0 else relative_embeddings)[:, slice_start_position:(slice_start_position + 2 * length - 1)]
-        else:
-            pad_length = max(length - (self.window_size + 1), 0)
-            slice_start_position = max((self.window_size + 1) - length, 0)
-
-            return (F.pad(relative_embeddings, convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]])) if pad_length > 0 else relative_embeddings)[:, slice_start_position:(slice_start_position + 2 * length - 1)]  
-
-    def _relative_position_to_absolute_position(self, x, onnx=False):
-        batch, heads, length, _ = x.size()
-
-        return (F.pad(F.pad(x, [0, 1, 0, 0, 0, 0, 0, 0]).view([batch, heads, length * 2 * length]), [0, length - 1, 0, 0, 0, 0]).view([batch, heads, length + 1, 2 * length - 1]) if onnx else F.pad(F.pad(x, convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, 1]])).view([batch, heads, length * 2 * length]), convert_pad_shape([[0, 0], [0, 0], [0, length - 1]])).view([batch, heads, length + 1, 2 * length - 1]))[:, :, :length, length - 1 :]
-
-    def _absolute_position_to_relative_position(self, x, onnx=False):
-        batch, heads, length, _ = x.size()
-
-        return (F.pad(F.pad(x, [0, length - 1, 0, 0, 0, 0, 0, 0]).view([batch, heads, length*length + length * (length - 1)]), [length, 0, 0, 0, 0, 0]).view([batch, heads, length, 2 * length]) if onnx else F.pad(F.pad(x, convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length - 1]])).view([batch, heads, length**2 + length * (length - 1)]), convert_pad_shape([[0, 0], [0, 0], [length, 0]])).view([batch, heads, length, 2 * length]))[:, :, :, 1:]
-
-    def _attention_bias_proximal(self, length):
-        r = torch.arange(length, dtype=torch.float32)
-
-        return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs((torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)))), 0), 0)
-
-class FFN(torch.nn.Module):
-    def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0.0, activation=None, causal=False, onnx=False):
-        super().__init__()
-        self.in_channels = in_channels
-        self.out_channels = out_channels
-        self.filter_channels = filter_channels
-        self.kernel_size = kernel_size
-        self.p_dropout = p_dropout
-        self.activation = activation
-        self.causal = causal
-        self.onnx = onnx
-        self.padding = self._causal_padding if causal else self._same_padding
-        self.conv_1 = torch.nn.Conv1d(in_channels, filter_channels, kernel_size)
-        self.conv_2 = torch.nn.Conv1d(filter_channels, out_channels, kernel_size)
-        self.drop = torch.nn.Dropout(p_dropout)
-
-    def forward(self, x, x_mask):
-        x = self.conv_1(self.padding(x * x_mask))
-
-        return self.conv_2(self.padding(self.drop(((x * torch.sigmoid(1.702 * x)) if self.activation == "gelu" else torch.relu(x))) * x_mask)) * x_mask
-
-    def _causal_padding(self, x):
-        if self.kernel_size == 1: return x
-
-        return F.pad(x, [self.kernel_size - 1, 0, 0, 0, 0, 0]) if self.onnx else F.pad(x, convert_pad_shape([[0, 0], [0, 0], [(self.kernel_size - 1), 0]]))
-
-    def _same_padding(self, x):
-        if self.kernel_size == 1: return x
-        
-        return F.pad(x, [(self.kernel_size - 1) // 2, self.kernel_size // 2, 0, 0, 0, 0]) if self.onnx else F.pad(x, convert_pad_shape([[0, 0], [0, 0], [((self.kernel_size - 1) // 2), (self.kernel_size // 2)]]))
-
-class Encoder(torch.nn.Module):
-    def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0.0, window_size=10, onnx=False, **kwargs):
-        super().__init__()
-        self.hidden_channels = hidden_channels
-        self.filter_channels = filter_channels
-        self.n_heads = n_heads
-        self.n_layers = n_layers
-        self.kernel_size = kernel_size
-        self.p_dropout = p_dropout
-        self.window_size = window_size
-        self.drop = torch.nn.Dropout(p_dropout)
-        self.attn_layers = torch.nn.ModuleList()
-        self.norm_layers_1 = torch.nn.ModuleList()
-        self.ffn_layers = torch.nn.ModuleList()
-        self.norm_layers_2 = torch.nn.ModuleList()
-
-        for _ in range(self.n_layers):
-            self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, window_size=window_size, onnx=onnx))
-            self.norm_layers_1.append(LayerNorm(hidden_channels, onnx=onnx))
-
-            self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, onnx=onnx))
-            self.norm_layers_2.append(LayerNorm(hidden_channels, onnx=onnx))
-
-    def forward(self, x, x_mask):
-        attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
-        x = x * x_mask
-        
-        for i in range(self.n_layers):
-            x = self.norm_layers_1[i](x + self.drop(self.attn_layers[i](x, x, attn_mask)))
-            x = self.norm_layers_2[i](x + self.drop(self.ffn_layers[i](x, x_mask)))
-
-        return x * x_mask
-
-class TextEncoder(torch.nn.Module):
-    def __init__(self, out_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, embedding_dim, f0=True, onnx=False):
-        super(TextEncoder, self).__init__()
-        self.out_channels = out_channels
-        self.hidden_channels = hidden_channels
-        self.filter_channels = filter_channels
-        self.n_heads = n_heads
-        self.n_layers = n_layers
-        self.kernel_size = kernel_size
-        self.p_dropout = float(p_dropout)
-        self.emb_phone = torch.nn.Linear(embedding_dim, hidden_channels)
-        self.lrelu = torch.nn.LeakyReLU(0.1, inplace=True)
-        if f0: self.emb_pitch = torch.nn.Embedding(256, hidden_channels)
-        self.encoder = Encoder(hidden_channels, filter_channels, n_heads, n_layers, kernel_size, float(p_dropout), onnx=onnx)
-        self.proj = torch.nn.Conv1d(hidden_channels, out_channels * 2, 1)
-
-    def forward(self, phone, pitch, lengths):
-        x = torch.transpose(self.lrelu(((self.emb_phone(phone) if pitch is None else (self.emb_phone(phone) + self.emb_pitch(pitch))) * math.sqrt(self.hidden_channels))), 1, -1) 
-        x_mask = torch.unsqueeze(sequence_mask(lengths, x.size(2)), 1).to(x.dtype)
-        m, logs = torch.split((self.proj(self.encoder(x * x_mask, x_mask)) * x_mask), self.out_channels, dim=1)
-
-        return m, logs, x_mask
-
-class PosteriorEncoder(torch.nn.Module):
-    def __init__(self, in_channels, out_channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0):
-        super(PosteriorEncoder, self).__init__()
-        self.in_channels = in_channels
-        self.out_channels = out_channels
-        self.hidden_channels = hidden_channels
-        self.kernel_size = kernel_size
-        self.dilation_rate = dilation_rate
-        self.n_layers = n_layers
-        self.gin_channels = gin_channels
-        self.pre = torch.nn.Conv1d(in_channels, hidden_channels, 1)
-        self.enc = WaveNet(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)
-        self.proj = torch.nn.Conv1d(hidden_channels, out_channels * 2, 1)
-
-    def forward(self, x, x_lengths, g = None):
-        x_mask = torch.unsqueeze(sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
-        m, logs = torch.split((self.proj(self.enc((self.pre(x) * x_mask), x_mask, g=g)) * x_mask), self.out_channels, dim=1)
-
-        return ((m + torch.randn_like(m) * torch.exp(logs)) * x_mask), m, logs, x_mask
-
-    def remove_weight_norm(self):
-        self.enc.remove_weight_norm()
-
-class Synthesizer(torch.nn.Module):
-    def __init__(self, spec_channels, segment_size, inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, spk_embed_dim, gin_channels, sr, use_f0, text_enc_hidden_dim=768, vocoder="Default", checkpointing=False, onnx=False, **kwargs):
-        super(Synthesizer, self).__init__()
-        self.spec_channels = spec_channels
-        self.inter_channels = inter_channels
-        self.hidden_channels = hidden_channels
-        self.filter_channels = filter_channels
-        self.n_heads = n_heads
-        self.n_layers = n_layers
-        self.kernel_size = kernel_size
-        self.p_dropout = float(p_dropout)
-        self.resblock_kernel_sizes = resblock_kernel_sizes
-        self.resblock_dilation_sizes = resblock_dilation_sizes
-        self.upsample_rates = upsample_rates
-        self.upsample_initial_channel = upsample_initial_channel
-        self.upsample_kernel_sizes = upsample_kernel_sizes
-        self.segment_size = segment_size
-        self.gin_channels = gin_channels
-        self.spk_embed_dim = spk_embed_dim
-        self.use_f0 = use_f0
-        self.enc_p = TextEncoder(inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, float(p_dropout), text_enc_hidden_dim, f0=use_f0, onnx=onnx)
-
-        if use_f0:
-            if vocoder == "RefineGAN": self.dec = RefineGANGenerator(sample_rate=sr, upsample_rates=upsample_rates, num_mels=inter_channels, checkpointing=checkpointing)
-            elif vocoder in ["MRF-HiFi-GAN", "MRF HiFi-GAN"]: self.dec = HiFiGANMRFGenerator(in_channel=inter_channels, upsample_initial_channel=upsample_initial_channel, upsample_rates=upsample_rates, upsample_kernel_sizes=upsample_kernel_sizes, resblock_kernel_sizes=resblock_kernel_sizes, resblock_dilations=resblock_dilation_sizes, gin_channels=gin_channels, sample_rate=sr, harmonic_num=8, checkpointing=checkpointing)
-            else: self.dec = GeneratorNSF(inter_channels, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels, sr=sr, checkpointing=checkpointing)
-        else: self.dec = Generator(inter_channels, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)
-
-        self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
-        self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels)
-        self.emb_g = torch.nn.Embedding(self.spk_embed_dim, gin_channels)
-
-    def remove_weight_norm(self):
-        self.dec.remove_weight_norm()
-        self.flow.remove_weight_norm()
-        self.enc_q.remove_weight_norm()
-
-    @torch.jit.ignore
-    def forward(self, phone, phone_lengths, pitch = None, pitchf = None, y = None, y_lengths = None, ds = None):
-        g = self.emb_g(ds).unsqueeze(-1)
-        m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
-
-        if y is not None:
-            z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
-            z_slice, ids_slice = rand_slice_segments(z, y_lengths, self.segment_size)
-
-            return (self.dec(z_slice, slice_segments(pitchf, ids_slice, self.segment_size, 2), g=g) if self.use_f0 else self.dec(z_slice, g=g)), ids_slice, x_mask, y_mask, (z, self.flow(z, y_mask, g=g), m_p, logs_p, m_q, logs_q)
-        else: return None, None, x_mask, None, (None, None, m_p, logs_p, None, None)
-
-    @torch.jit.export
-    def infer(self, phone, phone_lengths, pitch = None, nsff0 = None, sid = None, rate = None):
-        g = self.emb_g(sid).unsqueeze(-1)
-        m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
-        z_p = (m_p + torch.exp(logs_p) * torch.randn_like(m_p) * 0.66666) * x_mask
-
-        if rate is not None:
-            assert isinstance(rate, torch.Tensor)
-            head = int(z_p.shape[2] * (1.0 - rate.item()))
-            z_p = z_p[:, :, head:]
-            x_mask = x_mask[:, :, head:]
-            if self.use_f0: nsff0 = nsff0[:, head:]
-
-        if self.use_f0:
-            z = self.flow(z_p, x_mask, g=g, reverse=True)
-            o = self.dec(z * x_mask, nsff0, g=g)
-        else:
-            z = self.flow(z_p, x_mask, g=g, reverse=True)
-            o = self.dec(z * x_mask, g=g)
-
-        return o, x_mask, (z, z_p, m_p, logs_p)
-    
-class SynthesizerONNX(Synthesizer):
-    def __init__(self, spec_channels, segment_size, inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, spk_embed_dim, gin_channels, sr, use_f0, text_enc_hidden_dim=768, vocoder="Default", checkpointing=False, **kwargs):
-        super().__init__(spec_channels, segment_size, inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, spk_embed_dim, gin_channels, sr, use_f0, text_enc_hidden_dim, vocoder, checkpointing, True)
-        self.speaker_map = None
-
-    def remove_weight_norm(self):
-        self.dec.remove_weight_norm()
-        self.flow.remove_weight_norm()
-        self.enc_q.remove_weight_norm()
-
-    def construct_spkmixmap(self, n_speaker):
-        self.speaker_map = torch.zeros((n_speaker, 1, 1, self.gin_channels))
-
-        for i in range(n_speaker):
-            self.speaker_map[i] = self.emb_g(torch.LongTensor([[i]]))
-
-        self.speaker_map = self.speaker_map.unsqueeze(0)
-
-    def forward(self, phone, phone_lengths, g=None, rnd=None, pitch=None, nsff0=None, max_len=None):
-        g = self.emb_g(g).unsqueeze(-1)
-        m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
-        z_p = (m_p + torch.exp(logs_p) * rnd) * x_mask
-
-        return self.dec((self.flow(z_p, x_mask, g=g, reverse=True) * x_mask)[:, :, :max_len], nsff0, g=g) if self.use_f0 else self.dec((self.flow(z_p, x_mask, g=g, reverse=True) * x_mask)[:, :, :max_len], g=g)

main/library/architectures/demucs_separator.py

@@ -1,180 +0,0 @@
-import os
-import sys
-import yaml
-import torch
-
-import numpy as np
-from hashlib import sha256
-
-sys.path.append(os.getcwd())
-
-from main.configs.config import Config
-from main.library.uvr5_separator import spec_utils, common_separator
-from main.library.uvr5_separator.demucs import hdemucs, states, apply
-
-translations = Config().translations
-sys.path.insert(0, os.path.join(os.getcwd(), "main", "library", "uvr5_separator"))
-
-DEMUCS_4_SOURCE_MAPPER = {common_separator.CommonSeparator.BASS_STEM: 0, common_separator.CommonSeparator.DRUM_STEM: 1, common_separator.CommonSeparator.OTHER_STEM: 2, common_separator.CommonSeparator.VOCAL_STEM: 3}
-
-
-class DemucsSeparator(common_separator.CommonSeparator):
-    def __init__(self, common_config, arch_config):
-        super().__init__(config=common_config)
-        self.segment_size = arch_config.get("segment_size", "Default")
-        self.shifts = arch_config.get("shifts", 2)
-        self.overlap = arch_config.get("overlap", 0.25)
-        self.segments_enabled = arch_config.get("segments_enabled", True)
-        self.logger.debug(translations["demucs_info"].format(segment_size=self.segment_size, segments_enabled=self.segments_enabled))
-        self.logger.debug(translations["demucs_info_2"].format(shifts=self.shifts, overlap=self.overlap))
-        self.demucs_source_map = DEMUCS_4_SOURCE_MAPPER
-        self.audio_file_path = None
-        self.audio_file_base = None
-        self.demucs_model_instance = None
-        self.logger.info(translations["start_demucs"])
-
-    def separate(self, audio_file_path):
-        self.logger.debug(translations["start_separator"])
-        source = None
-        inst_source = {}
-        self.audio_file_path = audio_file_path
-        self.audio_file_base = os.path.splitext(os.path.basename(audio_file_path))[0]
-        self.logger.debug(translations["prepare_mix"])
-        mix = self.prepare_mix(self.audio_file_path)
-        self.logger.debug(translations["demix"].format(shape=mix.shape))
-        self.logger.debug(translations["cancel_mix"])
-        self.demucs_model_instance = hdemucs.HDemucs(sources=["drums", "bass", "other", "vocals"])
-        self.demucs_model_instance = get_demucs_model(name=os.path.splitext(os.path.basename(self.model_path))[0], repo=os.path.dirname(self.model_path))
-        self.demucs_model_instance = apply.demucs_segments(self.segment_size, self.demucs_model_instance)
-        self.demucs_model_instance.to(self.torch_device)
-        self.demucs_model_instance.eval()
-        self.logger.debug(translations["model_review"])
-        source = self.demix_demucs(mix)
-        del self.demucs_model_instance
-        self.clear_gpu_cache()
-        self.logger.debug(translations["del_gpu_cache_after_demix"])
-        output_files = []
-        self.logger.debug(translations["process_output_file"])
-
-        if isinstance(inst_source, np.ndarray):
-            self.logger.debug(translations["process_ver"])
-            inst_source[self.demucs_source_map[common_separator.CommonSeparator.VOCAL_STEM]] = spec_utils.reshape_sources(inst_source[self.demucs_source_map[common_separator.CommonSeparator.VOCAL_STEM]], source[self.demucs_source_map[common_separator.CommonSeparator.VOCAL_STEM]])
-            source = inst_source
-
-        if isinstance(source, np.ndarray):
-            source_length = len(source)
-            self.logger.debug(translations["source_length"].format(source_length=source_length))
-            self.logger.debug(translations["set_map"].format(part=source_length))
-
-            match source_length:
-                case 2: self.demucs_source_map = {common_separator.CommonSeparator.INST_STEM: 0, common_separator.CommonSeparator.VOCAL_STEM: 1}
-                case 6: self.demucs_source_map = {common_separator.CommonSeparator.BASS_STEM: 0, common_separator.CommonSeparator.DRUM_STEM: 1, common_separator.CommonSeparator.OTHER_STEM: 2, common_separator.CommonSeparator.VOCAL_STEM: 3, common_separator.CommonSeparator.GUITAR_STEM: 4, common_separator.CommonSeparator.PIANO_STEM: 5}
-                case _: self.demucs_source_map = DEMUCS_4_SOURCE_MAPPER
-
-        self.logger.debug(translations["process_all_part"])
-
-        for stem_name, stem_value in self.demucs_source_map.items():
-            if self.output_single_stem is not None:
-                if stem_name.lower() != self.output_single_stem.lower():
-                    self.logger.debug(translations["skip_part"].format(stem_name=stem_name, output_single_stem=self.output_single_stem))
-                    continue
-
-            stem_path = os.path.join(f"{self.audio_file_base}_({stem_name})_{self.model_name}.{self.output_format.lower()}")
-            self.final_process(stem_path, source[stem_value].T, stem_name)
-            output_files.append(stem_path)
-
-        return output_files
-
-    def demix_demucs(self, mix):
-        self.logger.debug(translations["starting_demix_demucs"])
-        processed = {}
-        mix = torch.tensor(mix, dtype=torch.float32)
-        ref = mix.mean(0)
-        mix = (mix - ref.mean()) / ref.std()
-        mix_infer = mix
-
-        with torch.no_grad():
-            self.logger.debug(translations["model_infer"])
-            sources = apply.apply_model(model=self.demucs_model_instance, mix=mix_infer[None], shifts=self.shifts, split=self.segments_enabled, overlap=self.overlap, static_shifts=1 if self.shifts == 0 else self.shifts, set_progress_bar=None, device=self.torch_device, progress=True)[0]
-
-        sources = (sources * ref.std() + ref.mean()).cpu().numpy()
-        sources[[0, 1]] = sources[[1, 0]]
-
-        processed[mix] = sources[:, :, 0:None].copy()
-        return np.concatenate([s[:, :, 0:None] for s in list(processed.values())], axis=-1)
-
-class LocalRepo:
-    def __init__(self, root):
-        self.root = root
-        self.scan()
-
-    def scan(self):
-        self._models, self._checksums = {}, {}
-        for filename in os.listdir(self.root):
-            filepath = os.path.join(self.root, filename)
-            if not os.path.isfile(filepath): continue
-
-            if os.path.splitext(filename)[1] == ".th":
-                stem = os.path.splitext(filename)[0]
-                
-                if "-" in stem:
-                    xp_sig, checksum = stem.split("-", 1)
-                    self._checksums[xp_sig] = checksum
-                else: xp_sig = stem
-
-                if xp_sig in self._models: raise RuntimeError(translations["del_all_but_one"].format(xp_sig=xp_sig))
-                self._models[xp_sig] = filepath
-
-    def has_model(self, sig):
-        return sig in self._models
-
-    def get_model(self, sig):
-        try:
-            file = self._models[sig]
-        except KeyError:
-            raise RuntimeError(translations["not_found_model_signature"].format(sig=sig))
-        
-        if sig in self._checksums: check_checksum(file, self._checksums[sig])
-        return states.load_model(file)
-
-class BagOnlyRepo:
-    def __init__(self, root, model_repo):
-        self.root = root
-        self.model_repo = model_repo
-        self.scan()
-
-    def scan(self):
-        self._bags = {}
-        for filename in os.listdir(self.root):
-            filepath = os.path.join(self.root, filename)
-
-            if os.path.isfile(filepath) and os.path.splitext(filename)[1] == ".yaml":
-                stem = os.path.splitext(filename)[0]
-                self._bags[stem] = filepath
-
-    def get_model(self, name):
-        try:
-            yaml_file = self._bags[name]
-        except KeyError:
-            raise RuntimeError(translations["name_not_pretrained"].format(name=name))
-        
-        with open(yaml_file, 'r') as f:
-            bag = yaml.safe_load(f)
-
-        return apply.BagOfModels([self.model_repo.get_model(sig) for sig in bag["models"]], bag.get("weights"), bag.get("segment"))
-
-def check_checksum(path, checksum):
-    sha = sha256()
-
-    with open(path, "rb") as file:
-        while 1:
-            buf = file.read(2**20)
-            if not buf: break
-            sha.update(buf)
-
-    actual_checksum = sha.hexdigest()[:len(checksum)]
-    if actual_checksum != checksum: raise RuntimeError(translations["invalid_checksum"].format(path=path, checksum=checksum, actual_checksum=actual_checksum))
-
-def get_demucs_model(name, repo = None):
-    model_repo = LocalRepo(repo)
-    return (model_repo.get_model(name) if model_repo.has_model(name) else BagOnlyRepo(repo, model_repo).get_model(name)).eval()

main/library/architectures/fairseq.py

@@ -1,1480 +0,0 @@
-import re
-import sys
-import math
-import uuid
-import torch
-import types
-import contextlib
-
-import numpy as np
-import torch.nn.functional as F
-
-from torch import nn
-from omegaconf import DictConfig, open_dict
-
-class Dictionary:
-    def __init__(self, *args, **kwargs):
-        pass
-
-fairseq = types.ModuleType("fairseq")
-fairseq_data = types.ModuleType("fairseq.data")
-fairseq_data_dictionary = types.ModuleType("fairseq.data.dictionary")
-fairseq_data_dictionary.Dictionary = Dictionary
-fairseq.data = fairseq_data
-fairseq_data.dictionary = fairseq_data_dictionary
-
-sys.modules["fairseq"] = fairseq
-sys.modules["fairseq.data"] = fairseq_data
-sys.modules["fairseq.data.dictionary"] = fairseq_data_dictionary
-
-def load_model(filename):
-    state = torch.load(filename, map_location="cpu")
-
-    model = HubertModel(HubertConfig(**state['cfg']['model']))
-    model.load_state_dict(state['model'], strict=False)
-
-    return [model], Model_Config(state["cfg"]), Model_Config(state["cfg"]["task"])
-
-def softmax(x, dim, onnx_trace = False):
-    return F.softmax(x.float(), dim=dim) if onnx_trace else F.softmax(x, dim=dim, dtype=torch.float32)
-
-def log_softmax(x, dim, onnx_trace = False):
-    return F.log_softmax(x.float(), dim=dim) if onnx_trace else F.log_softmax(x, dim=dim, dtype=torch.float32)
-
-def eval_str_dict(x, type=dict):
-    if x is None: return None
-    if isinstance(x, str): x = eval(x)
-    return x
-
-def with_incremental_state(cls):
-    cls.__bases__ = (FairseqIncrementalState,) + tuple(b for b in cls.__bases__ if b != FairseqIncrementalState)
-    return cls
-
-def quant_noise(module, p, block_size):
-    if p <= 0: return module
-    assert isinstance(module, (nn.Linear, nn.Embedding, nn.Conv2d))
-
-    is_conv = module.weight.ndim == 4
-    if not is_conv: assert (module.weight.size(1) % block_size == 0)
-    else:
-        if module.kernel_size == (1, 1): assert (module.in_channels % block_size == 0)
-        else:
-            k = module.kernel_size[0] * module.kernel_size[1]
-            assert k % block_size == 0
-
-    def _forward_pre_hook(mod, input):
-        if mod.training:
-            if not is_conv:
-                weight = mod.weight
-                in_features = weight.size(1)
-                out_features = weight.size(0)
-
-                mask = torch.zeros(in_features // block_size * out_features, device=weight.device)
-                mask.bernoulli_(p)
-                mask = mask.repeat_interleave(block_size, -1).view(-1, in_features)
-            else:
-                weight = mod.weight
-                in_channels = mod.in_channels
-                out_channels = mod.out_channels
-
-                if mod.kernel_size == (1, 1):
-                    mask = torch.zeros(int(in_channels // block_size * out_channels), device=weight.device)
-                    mask.bernoulli_(p)
-                    mask = mask.repeat_interleave(block_size, -1).view(-1, in_channels)
-                else:
-                    mask = torch.zeros(weight.size(0), weight.size(1), device=weight.device)
-                    mask.bernoulli_(p)
-                    mask = (mask.unsqueeze(2).unsqueeze(3).repeat(1, 1, mod.kernel_size[0], mod.kernel_size[1]))
-
-            mask = mask.to(torch.bool)
-            s = 1 / (1 - p)
-            mod.weight.data = s * weight.masked_fill(mask, 0)
-
-    module.register_forward_pre_hook(_forward_pre_hook)
-    return module
-
-class FairseqDropout(nn.Module):
-    def __init__(self, p, module_name=None):
-        super().__init__()
-        self.p = p
-        self.module_name = module_name
-        self.apply_during_inference = False
-
-    def forward(self, x, inplace = False):
-        return F.dropout(x, p=self.p, training=True, inplace=inplace) if self.p > 0 and (self.training or self.apply_during_inference) else x
-
-    def make_generation_fast_(self, name, retain_dropout = False, retain_dropout_modules = None, **kwargs):
-        if retain_dropout:
-            if (retain_dropout_modules is None or self.module_name in retain_dropout_modules): self.apply_during_inference = True
-
-class FairseqIncrementalState(object):
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.init_incremental_state()
-
-    def init_incremental_state(self):
-        self._incremental_state_id = str(uuid.uuid4())
-
-    def _get_full_incremental_state_key(self, key):
-        return "{}.{}".format(self._incremental_state_id, key)
-
-    def get_incremental_state(self, incremental_state, key):
-        full_key = self._get_full_incremental_state_key(key)
-        if incremental_state is None or full_key not in incremental_state: return None
-        return incremental_state[full_key]
-
-    def set_incremental_state(self, incremental_state, key, value):
-        if incremental_state is not None: incremental_state[self._get_full_incremental_state_key(key)] = value
-        return incremental_state
-
-class FairseqDecoder(nn.Module):
-    def __init__(self, dictionary):
-        super().__init__()
-        self.dictionary = dictionary
-        self.onnx_trace = False
-        self.adaptive_softmax = None
-
-    def forward(self, prev_output_tokens, encoder_out=None, **kwargs):
-        x, extra = self.extract_features(prev_output_tokens, encoder_out=encoder_out, **kwargs)
-        return self.output_layer(x), extra
-
-    def extract_features(self, prev_output_tokens, encoder_out=None, **kwargs):
-        pass
-
-    def output_layer(self, features, **kwargs):
-        pass
-
-    def get_normalized_probs(self, net_output, log_probs, sample = None):
-        return self.get_normalized_probs_scriptable(net_output, log_probs, sample)
-
-    def get_normalized_probs_scriptable(self, net_output, log_probs, sample = None):
-        if hasattr(self, "adaptive_softmax") and self.adaptive_softmax is not None:
-            if sample is not None:
-                assert "target" in sample
-                target = sample["target"]
-            else: target = None
-
-            out = self.adaptive_softmax.get_log_prob(net_output[0], target=target)
-            return out.exp_() if not log_probs else out
-
-        logits = net_output[0]
-        return log_softmax(logits, dim=-1, onnx_trace=self.onnx_trace) if log_probs else softmax(logits, dim=-1, onnx_trace=self.onnx_trace)
-
-    def max_positions(self):
-        return 1e6
-
-    def upgrade_state_dict_named(self, state_dict, name):
-        return state_dict
-
-    def prepare_for_onnx_export_(self):
-        self.onnx_trace = True
-
-@with_incremental_state
-class FairseqIncrementalDecoder(FairseqDecoder):
-    def __init__(self, dictionary):
-        super().__init__(dictionary)
-
-    def forward(self, prev_output_tokens, encoder_out=None, incremental_state=None, **kwargs):
-        pass
-
-    def extract_features(self, prev_output_tokens, encoder_out=None, incremental_state=None, **kwargs):
-        pass
-
-    def reorder_incremental_state(self, incremental_state, new_order):
-        pass
-
-    def reorder_incremental_state_scripting(self, incremental_state, new_order):
-        for module in self.modules():
-            if hasattr(module, "reorder_incremental_state"):
-                result = module.reorder_incremental_state(incremental_state, new_order)
-                if result is not None: incremental_state = result
-
-    def set_beam_size(self, beam_size):
-        if getattr(self, "_beam_size", -1) != beam_size:
-            seen = set()
-
-            def apply_set_beam_size(module):
-                if (module != self and hasattr(module, "set_beam_size") and module not in seen):
-                    seen.add(module)
-                    module.set_beam_size(beam_size)
-
-            self.apply(apply_set_beam_size)
-            self._beam_size = beam_size
-
-class MultiheadAttention(FairseqIncrementalDecoder):
-    def __init__(self, embed_dim, num_heads, kdim=None, vdim=None, dropout=0.0, bias=True, add_bias_kv=False, add_zero_attn=False, self_attention=False, encoder_decoder_attention=False, dictionary=None, q_noise=0.0, qn_block_size=8, xformers_att_config=None, xformers_blocksparse_layout=None, xformers_blocksparse_blocksize=16):
-        super().__init__(dictionary)
-        xformers_att_config = eval_str_dict(xformers_att_config)
-        self.use_xformers = xformers_att_config is not None
-        if self.use_xformers: raise ImportError
-        self.embed_dim = embed_dim
-        self.kdim = kdim if kdim is not None else embed_dim
-        self.vdim = vdim if vdim is not None else embed_dim
-        self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim
-        self.num_heads = num_heads
-        self.dropout_module = FairseqDropout(dropout, module_name=self.__class__.__name__)
-        self.head_dim = embed_dim // num_heads
-        assert (self.head_dim * num_heads == self.embed_dim)
-        self.scaling = self.head_dim**-0.5
-        self.self_attention = self_attention
-        self.encoder_decoder_attention = encoder_decoder_attention
-        assert not self.self_attention or self.qkv_same_dim
-        self.k_proj = quant_noise(nn.Linear(self.kdim, embed_dim, bias=bias), q_noise, qn_block_size)
-        self.v_proj = quant_noise(nn.Linear(self.vdim, embed_dim, bias=bias), q_noise, qn_block_size)
-        self.q_proj = quant_noise(nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size)
-        self.out_proj = quant_noise(nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size)
-        if add_bias_kv:
-            self.bias_k = nn.Parameter(torch.Tensor(1, 1, embed_dim))
-            self.bias_v = nn.Parameter(torch.Tensor(1, 1, embed_dim))
-        else: self.bias_k = self.bias_v = None
-        self.add_zero_attn = add_zero_attn
-        self.beam_size = 1
-        self.reset_parameters()
-        self.onnx_trace = False
-        self.skip_embed_dim_check = False
-        self.init_incremental_state()
-
-    def prepare_for_onnx_export_(self):
-        self.onnx_trace = True
-
-    def reset_parameters(self):
-        if self.qkv_same_dim:
-            nn.init.xavier_uniform_(self.k_proj.weight, gain=1 / math.sqrt(2))
-            nn.init.xavier_uniform_(self.v_proj.weight, gain=1 / math.sqrt(2))
-            nn.init.xavier_uniform_(self.q_proj.weight, gain=1 / math.sqrt(2))
-        else:
-            nn.init.xavier_uniform_(self.k_proj.weight)
-            nn.init.xavier_uniform_(self.v_proj.weight)
-            nn.init.xavier_uniform_(self.q_proj.weight)
-
-        nn.init.xavier_uniform_(self.out_proj.weight)
-
-        if self.out_proj.bias is not None: nn.init.constant_(self.out_proj.bias, 0.0)
-        if self.bias_k is not None: nn.init.xavier_normal_(self.bias_k)
-        if self.bias_v is not None: nn.init.xavier_normal_(self.bias_v)
-
-    def _get_reserve_head_index(self, num_heads_to_keep: int):
-        k_proj_heads_norm, q_proj_heads_norm, v_proj_heads_norm = [], [], []
-        for i in range(self.num_heads):
-            start_idx = i * self.head_dim
-            end_idx = (i + 1) * self.head_dim
-            k_proj_heads_norm.append(torch.sum(torch.abs(self.k_proj.weight[start_idx:end_idx])).tolist() + torch.sum(torch.abs(self.k_proj.bias[start_idx:end_idx])).tolist())
-            q_proj_heads_norm.append(torch.sum(torch.abs(self.q_proj.weight[start_idx:end_idx])).tolist() + torch.sum(torch.abs(self.q_proj.bias[start_idx:end_idx])).tolist())
-            v_proj_heads_norm.append(torch.sum(torch.abs(self.v_proj.weight[start_idx:end_idx])).tolist() + torch.sum(torch.abs(self.v_proj.bias[start_idx:end_idx])).tolist())
-
-        heads_norm = []
-        for i in range(self.num_heads):
-            heads_norm.append(k_proj_heads_norm[i] + q_proj_heads_norm[i] + v_proj_heads_norm[i])
-
-        sorted_head_index = sorted(range(self.num_heads), key=lambda k: heads_norm[k], reverse=True)
-        reserve_head_index = []
-        for i in range(num_heads_to_keep):
-            reserve_head_index.append((sorted_head_index[i] * self.head_dim, (sorted_head_index[i] + 1) * self.head_dim))
-        return reserve_head_index
-
-    def _adaptive_prune_heads(self, reserve_head_index):
-        new_q_weight, new_q_bias, new_k_weight, new_k_bias, new_v_weight, new_v_bias, new_out_proj_weight = [], [], [], [], [], [], []
-
-        for ele in reserve_head_index:
-            start_idx, end_idx = ele
-            new_q_weight.append(self.q_proj.weight[start_idx:end_idx])
-            new_q_bias.append(self.q_proj.bias[start_idx:end_idx])
-            new_k_weight.append(self.k_proj.weight[start_idx:end_idx])
-            new_k_bias.append(self.k_proj.bias[start_idx:end_idx])
-            new_v_weight.append(self.v_proj.weight[start_idx:end_idx])
-            new_v_bias.append(self.v_proj.bias[start_idx:end_idx])
-            new_out_proj_weight.append(self.out_proj.weight[:, start_idx:end_idx])
-
-        new_q_weight = torch.cat(new_q_weight).detach()
-        new_k_weight = torch.cat(new_k_weight).detach()
-        new_v_weight = torch.cat(new_v_weight).detach()
-        new_out_proj_weight = torch.cat(new_out_proj_weight, dim=-1).detach()
-        new_q_weight.requires_grad = True
-        new_k_weight.requires_grad = True
-        new_v_weight.requires_grad = True
-        new_out_proj_weight.requires_grad = True
-        new_q_bias = torch.cat(new_q_bias).detach()
-        new_q_bias.requires_grad = True
-        new_k_bias = torch.cat(new_k_bias).detach()
-        new_k_bias.requires_grad = True
-        new_v_bias = torch.cat(new_v_bias).detach()
-        new_v_bias.requires_grad = True
-
-        self.q_proj.weight = nn.Parameter(new_q_weight)
-        self.q_proj.bias = nn.Parameter(new_q_bias)
-        self.k_proj.weight = nn.Parameter(new_k_weight)
-        self.k_proj.bias = nn.Parameter(new_k_bias)
-        self.v_proj.weight = nn.Parameter(new_v_weight)
-        self.v_proj.bias = nn.Parameter(new_v_bias)
-        self.out_proj.weight = nn.Parameter(new_out_proj_weight)
-        self.num_heads = len(reserve_head_index)
-        self.embed_dim = self.head_dim * self.num_heads
-        self.q_proj.out_features = self.embed_dim
-        self.k_proj.out_features = self.embed_dim
-        self.v_proj.out_features = self.embed_dim
-
-    def _set_skip_embed_dim_check(self):
-        self.skip_embed_dim_check = True
-
-    def _pad_masks(self, key_padding_mask, attn_mask):
-        if attn_mask is not None:
-            shape = attn_mask.size()[:-1] + torch.Size([1])
-            attn_mask = torch.cat([attn_mask, attn_mask.new_zeros(shape)], dim=-1)
-
-        if key_padding_mask is not None:
-            shape = key_padding_mask.size()[:-1] + torch.Size([1])
-            key_padding_mask = torch.cat([key_padding_mask, key_padding_mask.new_zeros(shape)], dim=-1)
-
-        return key_padding_mask, attn_mask
-
-    def _add_bias(self, k, v, key_padding_mask, attn_mask, bsz):
-        assert self.bias_k is not None or self.bias_v is not None
-        key_padding_mask, attn_mask = self._pad_masks(key_padding_mask=key_padding_mask, attn_mask=attn_mask)
-        return torch.cat([k, self.bias_k.repeat(1, bsz, 1)]), torch.cat([v, self.bias_v.repeat(1, bsz, 1)]), key_padding_mask, attn_mask
-
-    def _append_zero_attn(self, k, v, key_padding_mask, attn_mask):
-        zero_attn_shape = k.size()[:-2] + torch.Size([1]) + k.size()[-1:]
-        key_padding_mask, attn_mask = self._pad_masks(key_padding_mask=key_padding_mask, attn_mask=attn_mask)
-        return torch.cat([k, torch.zeros(zero_attn_shape, dtype=k.dtype, device=k.device)], dim=-2), torch.cat([v, torch.zeros(zero_attn_shape, dtype=v.dtype, device=v.device)], dim=-2), key_padding_mask, attn_mask
-
-    def forward(self, query, key, value, key_padding_mask = None, incremental_state = None, need_weights = True, static_kv = False, attn_mask = None, before_softmax = False, need_head_weights = False):
-        if need_head_weights: need_weights = True
-        is_tpu = query.device.type == "xla"
-        tgt_len, bsz, embed_dim = query.size()
-        src_len = tgt_len
-        
-        if not self.skip_embed_dim_check: assert (embed_dim == self.embed_dim)
-        assert list(query.size()) == [tgt_len, bsz, embed_dim]
-
-        if key is not None:
-            src_len, key_bsz, _ = key.size()
-            if not torch.jit.is_scripting():
-                assert value is not None
-                assert src_len, key_bsz == value.shape[:2]
-
-        if (not self.onnx_trace and not is_tpu and incremental_state is None and not static_kv and not torch.jit.is_scripting() and not self.skip_embed_dim_check):
-            assert key is not None and value is not None
-            return F.multi_head_attention_forward(query, key, value, self.embed_dim, self.num_heads, torch.empty([0]), torch.cat((self.q_proj.bias, self.k_proj.bias, self.v_proj.bias)), self.bias_k, self.bias_v, self.add_zero_attn, self.dropout_module.p, self.out_proj.weight, self.out_proj.bias, self.training or self.dropout_module.apply_during_inference, key_padding_mask.bool() if key_padding_mask is not None else None, need_weights, attn_mask, use_separate_proj_weight=True, q_proj_weight=self.q_proj.weight, k_proj_weight=self.k_proj.weight, v_proj_weight=self.v_proj.weight)
-
-        if incremental_state is not None:
-            saved_state = self._get_input_buffer(incremental_state)
-            if saved_state is not None and "prev_key" in saved_state:
-                if static_kv:
-                    assert self.encoder_decoder_attention and not self.self_attention
-                    key = value = None
-        else: saved_state = None
-
-        if self.self_attention:
-            q = self.q_proj(query)
-            k = self.k_proj(query)
-            v = self.v_proj(query)
-        elif self.encoder_decoder_attention:
-            q = self.q_proj(query)
-            if key is None:
-                assert value is None
-                k = v = None
-            else:
-                if self.beam_size > 1 and bsz == key.size(1):
-                    key = key.view(key.size(0), -1, self.beam_size, key.size(2))[:, :, 0, :]
-                    if key_padding_mask is not None: key_padding_mask = key_padding_mask.view(-1, self.beam_size, key_padding_mask.size(1))[:, 0, :]
-                k = self.k_proj(key)
-                v = self.v_proj(key)
-        else:
-            assert key is not None and value is not None
-            q = self.q_proj(query)
-            k = self.k_proj(key)
-            v = self.v_proj(value)
-
-        q *= self.scaling
-
-        if self.bias_k is not None:
-            assert self.bias_v is not None
-            k, v, attn_mask, key_padding_mask = self._add_bias(k, v, attn_mask, key_padding_mask, bsz)
-
-        q = (q.contiguous().view(tgt_len, bsz * self.num_heads, self.head_dim).transpose(0, 1))
-        kv_bsz = bsz 
-
-        if k is not None:
-            kv_bsz = k.size(1)
-            k = (k.contiguous().view(-1, kv_bsz * self.num_heads, self.head_dim).transpose(0, 1))
-
-        if v is not None: v = (v.contiguous().view(-1, kv_bsz * self.num_heads, self.head_dim).transpose(0, 1))
-
-        if saved_state is not None:
-            if "prev_key" in saved_state:
-                _prev_key = saved_state["prev_key"]
-                assert _prev_key is not None
-
-                kv_bsz = _prev_key.size(0)
-                prev_key = _prev_key.view(kv_bsz * self.num_heads, -1, self.head_dim)
-
-                if static_kv: k = prev_key
-                else:
-                    assert k is not None
-                    k = torch.cat([prev_key, k], dim=1)
-                src_len = k.size(1)
-
-            if "prev_value" in saved_state:
-                _prev_value = saved_state["prev_value"]
-                assert _prev_value is not None or kv_bsz == _prev_value.size(0)
-                prev_value = _prev_value.view(kv_bsz * self.num_heads, -1, self.head_dim)
-
-                if static_kv: v = prev_value
-                else:
-                    assert v is not None
-                    v = torch.cat([prev_value, v], dim=1)
-
-            prev_key_padding_mask = None
-            if "prev_key_padding_mask" in saved_state: prev_key_padding_mask = saved_state["prev_key_padding_mask"]
-
-            assert k is not None and v is not None
-            key_padding_mask = MultiheadAttention._append_prev_key_padding_mask(key_padding_mask=key_padding_mask, prev_key_padding_mask=prev_key_padding_mask, batch_size=kv_bsz, src_len=k.size(1), static_kv=static_kv)
-
-            saved_state["prev_key"] = k.view(kv_bsz, self.num_heads, -1, self.head_dim)
-            saved_state["prev_value"] = v.view(kv_bsz, self.num_heads, -1, self.head_dim)
-            saved_state["prev_key_padding_mask"] = key_padding_mask
-
-            assert incremental_state is not None
-            incremental_state = self._set_input_buffer(incremental_state, saved_state)
-
-        assert k is not None
-        assert k.size(1) == src_len
-
-        if key_padding_mask is not None and key_padding_mask.dim() == 0: key_padding_mask = None
-
-        if key_padding_mask is not None:
-            assert key_padding_mask.size(0) == kv_bsz
-            assert key_padding_mask.size(1) == src_len
-
-        if self.add_zero_attn:
-            assert v is not None
-            src_len += 1
-            k, v, key_padding_mask, attn_mask = self._append_zero_attn(k=k, v=v, key_padding_mask=key_padding_mask, attn_mask=attn_mask)
-
-        if self.encoder_decoder_attention and bsz != kv_bsz:
-            attn_weights = torch.einsum("bxhtd,bhsd->bxhts", q.view((kv_bsz, -1, self.num_heads) + q.size()[1:]), k.view((kv_bsz, self.num_heads) + k.size()[1:]))
-            attn_weights = attn_weights.reshape((-1,) + attn_weights.size()[-2:])
-        else: attn_weights = torch.bmm(q, k.transpose(1, 2))
-
-        attn_weights = self.apply_sparse_mask(attn_weights, tgt_len, src_len, bsz)
-        assert list(attn_weights.size()) == [bsz * self.num_heads, tgt_len, src_len]
-
-        if attn_mask is not None:
-            attn_mask = attn_mask.unsqueeze(0)
-            if self.onnx_trace: attn_mask = attn_mask.repeat(attn_weights.size(0), 1, 1)
-            attn_weights += attn_mask
-
-        if key_padding_mask is not None:
-            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
-            attn_weights = attn_weights.view(kv_bsz, -1, self.num_heads, tgt_len, src_len).masked_fill(key_padding_mask.unsqueeze(1).unsqueeze(2).unsqueeze(3).to(torch.bool), float("-inf")) if not is_tpu else attn_weights.transpose(0, 2).masked_fill(key_padding_mask, float("-inf")).transpose(0, 2)
-            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
-
-        if before_softmax: return attn_weights, v
-
-        attn_weights_float = softmax(attn_weights, dim=-1, onnx_trace=self.onnx_trace)
-        attn_weights = attn_weights_float.type_as(attn_weights)
-        attn_probs = self.dropout_module(attn_weights)
-
-        assert v is not None
-        attn = None
-
-        if self.encoder_decoder_attention and bsz != kv_bsz:
-            attn = torch.einsum("bxhts,bhsd->bxhtd", attn_probs.view((kv_bsz, -1, self.num_heads) + attn_probs.size()[1:]), v.view((kv_bsz, self.num_heads) + v.size()[1:]))
-            attn = attn.reshape((-1,) + attn.size()[-2:])
-        else: attn = torch.bmm(attn_probs, v)
-        assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim]
-
-        if self.onnx_trace and attn.size(1) == 1: attn = attn.contiguous().view(tgt_len, bsz, self.embed_dim)
-        else: attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, self.embed_dim)
-
-        attn = self.out_proj(attn)
-        attn_weights = None
-
-        if need_weights:
-            attn_weights = attn_weights_float.view(bsz, self.num_heads, tgt_len, src_len).transpose(1, 0)
-            if not need_head_weights: attn_weights = attn_weights.mean(dim=0)
-
-        return attn, attn_weights
-
-    @staticmethod
-    def _append_prev_key_padding_mask(key_padding_mask, prev_key_padding_mask, batch_size, src_len, static_kv):
-        if prev_key_padding_mask is not None and static_kv: new_key_padding_mask = prev_key_padding_mask
-        elif prev_key_padding_mask is not None and key_padding_mask is not None: new_key_padding_mask = torch.cat([prev_key_padding_mask.float(), key_padding_mask.float()], dim=1)
-        elif prev_key_padding_mask is not None:
-            if src_len > prev_key_padding_mask.size(1):
-                filler = torch.zeros((batch_size, src_len - prev_key_padding_mask.size(1)), device=prev_key_padding_mask.device)
-                new_key_padding_mask = torch.cat([prev_key_padding_mask.float(), filler.float()], dim=1)
-            else: new_key_padding_mask = prev_key_padding_mask.float()
-        elif key_padding_mask is not None:
-            if src_len > key_padding_mask.size(1):
-                filler = torch.zeros((batch_size, src_len - key_padding_mask.size(1)), device=key_padding_mask.device)
-                new_key_padding_mask = torch.cat([filler.float(), key_padding_mask.float()], dim=1)
-            else: new_key_padding_mask = key_padding_mask.float()
-        else: new_key_padding_mask = prev_key_padding_mask
-        return new_key_padding_mask
-
-    @torch.jit.export
-    def reorder_incremental_state(self, incremental_state, new_order):
-        input_buffer = self._get_input_buffer(incremental_state)
-        if input_buffer is not None:
-            for k in input_buffer.keys():
-                input_buffer_k = input_buffer[k]
-                if input_buffer_k is not None:
-                    if self.encoder_decoder_attention:
-                        if input_buffer_k.size(0) * self.beam_size == new_order.size(0): return incremental_state
-                        elif self.beam_size > 1: input_buffer[k] = input_buffer_k.index_select(0, new_order.reshape(-1, self.beam_size)[:, 0] // self.beam_size)
-                        else: input_buffer[k] = input_buffer_k.index_select(0, new_order)
-                    else: input_buffer[k] = input_buffer_k.index_select(0, new_order)
-            incremental_state = self._set_input_buffer(incremental_state, input_buffer)
-        return incremental_state
-
-    def set_beam_size(self, beam_size):
-        self.beam_size = beam_size
-
-    def _get_input_buffer(self, incremental_state):
-        result = self.get_incremental_state(incremental_state, "attn_state")
-        if result is not None: return result
-        else: return {}
-
-    def _set_input_buffer(self, incremental_state, buffer):
-        return self.set_incremental_state(incremental_state, "attn_state", buffer)
-
-    def apply_sparse_mask(self, attn_weights, tgt_len: int, src_len: int, bsz: int):
-        return attn_weights
-
-    def upgrade_state_dict_named(self, state_dict, name):
-        prefix = name + "." if name != "" else ""
-        items_to_add = {}
-        keys_to_remove = []
-        for k in state_dict.keys():
-            if k.endswith(prefix + "in_proj_weight"):
-                dim = int(state_dict[k].shape[0] / 3)
-                items_to_add[prefix + "q_proj.weight"] = state_dict[k][:dim]
-                items_to_add[prefix + "k_proj.weight"] = state_dict[k][dim : 2 * dim]
-                items_to_add[prefix + "v_proj.weight"] = state_dict[k][2 * dim :]
-                keys_to_remove.append(k)
-                k_bias = prefix + "in_proj_bias"
-                if k_bias in state_dict.keys():
-                    dim = int(state_dict[k].shape[0] / 3)
-                    items_to_add[prefix + "q_proj.bias"] = state_dict[k_bias][:dim]
-                    items_to_add[prefix + "k_proj.bias"] = state_dict[k_bias][dim : 2 * dim]
-                    items_to_add[prefix + "v_proj.bias"] = state_dict[k_bias][2 * dim :]
-                    keys_to_remove.append(prefix + "in_proj_bias")
-
-        for k in keys_to_remove:
-            del state_dict[k]
-
-        for key, value in items_to_add.items():
-            state_dict[key] = value
-
-def init_bert_params(module):
-    def normal_(data):
-        data.copy_(data.cpu().normal_(mean=0.0, std=0.02).to(data.device))
-
-    if isinstance(module, nn.Linear):
-        normal_(module.weight.data)
-        if module.bias is not None: module.bias.data.zero_()
-    if isinstance(module, nn.Embedding):
-        normal_(module.weight.data)
-        if module.padding_idx is not None: module.weight.data[module.padding_idx].zero_()
-    if isinstance(module, MultiheadAttention):
-        normal_(module.q_proj.weight.data)
-        normal_(module.k_proj.weight.data)
-        normal_(module.v_proj.weight.data)
-
-def make_conv_pos(e, k, g):
-    pos_conv = nn.Conv1d(e, e, kernel_size=k, padding=k // 2, groups=g)
-    dropout = 0
-
-    nn.init.normal_(pos_conv.weight, mean=0, std=math.sqrt((4 * (1.0 - dropout)) / (k * e)))
-    nn.init.constant_(pos_conv.bias, 0)
-
-    return nn.Sequential(nn.utils.parametrizations.weight_norm(pos_conv, name="weight", dim=2), SamePad(k), nn.GELU())
-
-def is_xla_tensor(tensor):
-    return torch.is_tensor(tensor) and tensor.device.type == "xla"
-
-def index_put(tensor, indices, value):
-    if is_xla_tensor(tensor):
-        for _ in range(indices.dim(), tensor.dim()): 
-            indices = indices.unsqueeze(-1)
-
-        if indices.size(-1) < tensor.size(-1): indices = indices.expand_as(tensor)
-        tensor = torch.mul(tensor, ~indices) + torch.mul(value, indices)
-    else: tensor[indices] = value
-
-    return tensor
-
-def pad_to_multiple(x, multiple, dim=-1, value=0):
-    if x is None: return None, 0
-    tsz = x.size(dim)
-    m = tsz / multiple
-    remainder = math.ceil(m) * multiple - tsz
-    if m.is_integer(): return x, 0
-    return F.pad(x, (*((0,) * (-1 - dim) * 2), 0, remainder), value=value), remainder
-
-def compute_mask_indices(shape, padding_mask, mask_prob, mask_length, mask_type = "static", mask_other = 0.0, min_masks = 0, no_overlap = False, min_space = 0, require_same_masks = True, mask_dropout = 0.0, add_masks = False, seed = None, epoch = None, indices = None, idc_select_ver = 1, num_mask_ver = 2):
-    bsz, all_sz = shape
-    mask = np.full((bsz, all_sz), False)
-
-    if num_mask_ver == 1: all_num_mask = max(min_masks, int(mask_prob * all_sz / float(mask_length) + np.random.rand()))
-    mask_idcs = []
-
-    for i in range(bsz):
-        seed_i = int(hash((seed, epoch, indices[i].item())) % 1e6) if seed is not None and epoch is not None and indices is not None else None
-        rng = np.random.default_rng(seed_i)
-
-        if padding_mask is not None:
-            sz = all_sz - padding_mask[i].long().sum().item()
-            assert sz >= 0, sz
-        else: sz = all_sz
-
-        if num_mask_ver == 1: num_mask = max(min_masks, int(mask_prob * sz / float(mask_length) + np.random.rand())) if padding_mask is not None else all_num_mask
-        elif num_mask_ver == 2: num_mask = max(min_masks, int(mask_prob * sz / float(mask_length) + rng.random()))
-        else: raise ValueError
-
-        if mask_type == "static": lengths = np.full(num_mask, mask_length)
-        elif mask_type == "uniform": lengths = rng.randint(mask_other, mask_length * 2 + 1, size=num_mask)
-        elif mask_type == "normal": lengths = [max(1, int(round(x))) for x in rng.normal(mask_length, mask_other, size=num_mask)]
-        elif mask_type == "poisson": lengths = [int(round(x)) for x in rng.poisson(mask_length, size=num_mask)]
-        else: raise Exception
-
-        if sum(lengths) == 0:
-            if mask_type == "static": raise ValueError
-            else: lengths = [min(mask_length, sz - 1)]
-
-        if no_overlap:
-            mask_idc = []
-
-            def arrange(s, e, length, keep_length):
-                span_start = rng.randint(s, e - length)
-                mask_idc.extend(span_start + i for i in range(length))
-                new_parts = []
-
-                if span_start - s - min_space >= keep_length: new_parts.append((s, span_start - min_space + 1))
-                if e - span_start - length - min_space > keep_length: new_parts.append((span_start + length + min_space, e))
-
-                return new_parts
-
-            parts = [(0, sz)]
-            min_length = min(lengths)
-
-            for length in sorted(lengths, reverse=True):
-                lens = np.fromiter((e - s if e - s >= length + min_space else 0 for s, e in parts), np.int32)
-                l_sum = np.sum(lens)
-                if l_sum == 0: break
-                s, e = parts.pop(rng.choice(len(parts), p=lens / np.sum(lens)))
-                parts.extend(arrange(s, e, length, min_length))
-            mask_idc = np.asarray(mask_idc)
-        else:
-            if idc_select_ver == 1:
-                min_len = min(lengths)
-                if sz - min_len <= num_mask: min_len = sz - num_mask - 1
-                mask_idc = rng.choice(sz - min_len, num_mask, replace=False)
-            elif idc_select_ver == 2: mask_idc = rng.choice(sz, num_mask, replace=False)
-            else: raise ValueError
-
-            mask_idc = np.asarray([mask_idc[j] + offset for j in range(len(mask_idc)) for offset in range(lengths[j])])
-
-        mask_idc = np.unique(mask_idc[mask_idc < sz])
-        if len(mask_idc) >= sz: raise ValueError
-        mask_idcs.append(mask_idc)
-
-    target_len = None
-    if require_same_masks: target_len = max([len(m) for m in mask_idcs]) if add_masks else min([len(m) for m in mask_idcs])
-
-    for i, mask_idc in enumerate(mask_idcs):
-        if target_len is not None and len(mask_idc) > target_len: mask_idc = rng.choice(mask_idc, target_len, replace=False)
-        mask[i, mask_idc] = True
-
-        if target_len is not None and len(mask_idc) < target_len:
-            to_mask = rng.choice(np.flatnonzero(~mask[i]), target_len - len(mask_idc), replace=False)
-            mask[i, to_mask] = True
-
-        if mask_dropout > 0:
-            masked = np.flatnonzero(mask[i])
-            mask[i, rng.choice(masked, np.rint(len(masked) * mask_dropout).astype(int), replace=False)] = False
-
-    return mask
-
-def LayerNorm(normalized_shape, eps=1e-5, elementwise_affine=True):
-    return nn.LayerNorm(normalized_shape, eps, elementwise_affine)
-
-def prune_state_dict(state_dict, model_cfg):
-    arch = None
-    if model_cfg is not None: arch = (model_cfg._name if isinstance(model_cfg, DictConfig) else getattr(model_cfg, "arch", None))
-
-    if not model_cfg or arch is None or arch == "ptt_transformer": return state_dict
-
-    encoder_layers_to_keep = getattr(model_cfg, "encoder_layers_to_keep", None)
-    decoder_layers_to_keep = getattr(model_cfg, "decoder_layers_to_keep", None)
-
-    if not encoder_layers_to_keep and not decoder_layers_to_keep: return state_dict
-
-    def create_pruning_pass(layers_to_keep, layer_name):
-        keep_layers = sorted(int(layer_string) for layer_string in layers_to_keep.split(","))
-        mapping_dict = {}
-        for i in range(len(keep_layers)):
-            mapping_dict[str(keep_layers[i])] = str(i)
-
-        return {"substitution_regex": re.compile(r"^{layer}.*\.layers\.(\d+)".format(layer=layer_name)), "mapping_dict": mapping_dict}
-
-    pruning_passes = []
-    new_state_dict = {}
-
-    if encoder_layers_to_keep: pruning_passes.append(create_pruning_pass(encoder_layers_to_keep, "encoder"))
-    if decoder_layers_to_keep: pruning_passes.append(create_pruning_pass(decoder_layers_to_keep, "decoder"))
-
-    for layer_name in state_dict.keys():
-        match = re.search(r"\.layers\.(\d+)\.", layer_name)
-        if not match:
-            new_state_dict[layer_name] = state_dict[layer_name]
-            continue
-
-        original_layer_number = match.group(1)
-        for pruning_pass in pruning_passes:
-            if original_layer_number in pruning_pass["mapping_dict"] and pruning_pass["substitution_regex"].search(layer_name):
-                substitution_match = pruning_pass["substitution_regex"].search(layer_name)
-                new_state_dict[(layer_name[: substitution_match.start(1)] + pruning_pass["mapping_dict"][original_layer_number] + layer_name[substitution_match.end(1) :])] = state_dict[layer_name]
-
-    with open_dict(model_cfg) if isinstance(model_cfg, DictConfig) else contextlib.ExitStack():
-        if hasattr(model_cfg, "encoder_layers_to_keep"): model_cfg.encoder_layers_to_keep = None
-        if hasattr(model_cfg, "decoder_layers_to_keep"): model_cfg.decoder_layers_to_keep = None
-
-    return new_state_dict
-
-def relu_squared(x):
-    return F.relu(x).pow(2)
-
-def get_activation_fn(activation):
-    def gelu(x):
-        return nn.functional.gelu(x.float()).type_as(x)
-    
-    def gelu_accurate(x):
-        if not hasattr(gelu_accurate, "_a"):
-            gelu_accurate._a = math.sqrt(2 / math.pi)
-            return (0.5 * x * (1 + torch.tanh(gelu_accurate._a * (x + 0.044715 * torch.pow(x, 3)))))
-
-    if activation == "relu": return F.relu
-    elif activation == "relu_squared": return relu_squared
-    elif activation == "gelu": return gelu
-    elif activation == "gelu_fast": return gelu_accurate
-    elif activation == "gelu_accurate": return gelu_accurate
-    elif activation == "tanh": return torch.tanh
-    elif activation == "linear": return lambda x: x
-    elif activation == "swish": return nn.SiLU
-    else: raise RuntimeError
-
-class SamePad(nn.Module):
-    def __init__(self, kernel_size, causal=False):
-        super().__init__()
-        if causal: self.remove = kernel_size - 1
-        else: self.remove = 1 if kernel_size % 2 == 0 else 0
-
-    def forward(self, x):
-        if self.remove > 0: x = x[:, :, : -self.remove]
-        return x
-
-class TransformerSentenceEncoderLayer(nn.Module):
-    def __init__(self, embedding_dim = 768, ffn_embedding_dim = 3072, num_attention_heads = 8, dropout = 0.1, attention_dropout = 0.1, activation_dropout = 0.1, activation_fn = "relu", layer_norm_first = False):
-        super().__init__()
-        self.embedding_dim = embedding_dim
-        self.dropout = dropout
-        self.activation_dropout = activation_dropout
-        self.activation_fn = get_activation_fn(activation_fn)
-        self.self_attn = MultiheadAttention(self.embedding_dim, num_attention_heads, dropout=attention_dropout, self_attention=True)
-        self.dropout1 = nn.Dropout(dropout)
-        self.dropout2 = nn.Dropout(self.activation_dropout)
-        self.dropout3 = nn.Dropout(dropout)
-        self.layer_norm_first = layer_norm_first
-        self.self_attn_layer_norm = LayerNorm(self.embedding_dim)
-        self.fc1 = nn.Linear(self.embedding_dim, ffn_embedding_dim)
-        self.fc2 = nn.Linear(ffn_embedding_dim, self.embedding_dim)
-        self.final_layer_norm = LayerNorm(self.embedding_dim)
-
-    def forward(self, x, self_attn_mask=None, self_attn_padding_mask=None, need_weights=False, att_args=None):
-        residual = x
-
-        if self.layer_norm_first:
-            x = self.self_attn_layer_norm(x)
-            x, attn = self.self_attn(query=x, key=x, value=x, key_padding_mask=self_attn_padding_mask, attn_mask=self_attn_mask, need_weights=False)
-            x = residual + self.dropout1(x)
-            residual = x
-            x = self.fc2(self.dropout2(self.activation_fn(self.fc1(self.final_layer_norm(x)))))
-            layer_result = x
-            x = residual + self.dropout3(x)
-        else:
-            x, attn = self.self_attn(query=x, key=x, value=x, key_padding_mask=self_attn_padding_mask, need_weights=False)
-            x = self.self_attn_layer_norm(residual + self.dropout1(x))
-            residual = x
-            x = self.fc2(self.dropout2(self.activation_fn(self.fc1(x))))
-            layer_result = x
-            x = self.final_layer_norm(residual + self.dropout3(x))
-
-        return x, (attn, layer_result)
-
-class AdapterFast(nn.Module):
-    def __init__(self, adapter_num, input_dim, hidden_dim, act_fn):
-        super().__init__()
-        self.adapter_num = adapter_num
-        self.input_dim = input_dim
-        self.hidden_dim = hidden_dim
-        self.W_a = nn.Parameter(torch.empty(adapter_num, hidden_dim, input_dim))
-        self.W_b = nn.Parameter(torch.empty(adapter_num, input_dim, hidden_dim))
-        self.b_a = nn.Parameter(torch.empty(adapter_num, hidden_dim))
-        self.b_b = nn.Parameter(torch.empty(adapter_num, input_dim))
-        self.ln_W = nn.Parameter(torch.empty(adapter_num, input_dim))
-        self.ln_b = nn.Parameter(torch.empty(adapter_num, input_dim))
-        self.act_fn = nn.Identity()
-        if act_fn == "relu": self.act_fn = nn.ReLU()
-        elif act_fn == "gelu": self.act_fn = nn.GELU()
-        elif act_fn == "selu": self.act_fn = nn.SELU()
-        else: raise ValueError
-
-        self.input_dim = input_dim
-        self.reset_parameters()
-
-    def reset_parameters(self):
-        for ii in range(self.adapter_num):
-            nn.init.kaiming_uniform_(self.W_a[ii], a=math.sqrt(5))
-            nn.init.kaiming_uniform_(self.W_b[ii], a=math.sqrt(5))
-            fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.W_a[ii])
-            bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0
-            nn.init.uniform_(self.b_a[ii], -bound, bound)
-            fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.W_b[ii])
-            bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0
-            nn.init.uniform_(self.b_b[ii], -bound, bound)
-        
-        nn.init.ones_(self.ln_W)
-        nn.init.zeros_(self.ln_b)
-
-    def forward(self, x, adapter_id):
-        ii = adapter_id
-        return F.linear(self.act_fn(F.linear(F.layer_norm(x, (self.input_dim, ), self.ln_W[ii], self.ln_b[ii]), self.W_a[ii], self.b_a[ii])), self.W_b[ii], self.b_b[ii])
-    
-    def extra_repr(self):
-        return ('adapter={}, input_dim={}, hidden_dim={}'.format(self.adapter_num, self.input_dim, self.hidden_dim))
-
-class FeedForwardModule(nn.Module):
-    def __init__(self, input_feat, hidden_units, dropout1, dropout2, activation_fn="swish", bias=True):
-        super(FeedForwardModule, self).__init__()
-        self.layer_norm = LayerNorm(input_feat)
-        self.w_1 = nn.Linear(input_feat, hidden_units, bias=bias)
-        self.w_2 = nn.Linear(hidden_units, input_feat, bias=bias)
-        self.dropout1 = nn.Dropout(dropout1)
-        self.dropout2 = nn.Dropout(dropout2)
-        self.activation = get_activation_fn(activation_fn)(hidden_units)
-
-    def forward(self, x):
-        return self.dropout2(self.w_2(self.dropout1(self.activation(self.w_1(self.layer_norm(x))))))
-
-class ConvolutionModule(nn.Module):
-    def __init__(self, embed_dim, channels, depthwise_kernel_size, dropout, activation_fn="swish", bias=False, export=False):
-        super(ConvolutionModule, self).__init__()
-        assert (depthwise_kernel_size - 1) % 2 == 0
-        self.layer_norm = LayerNorm(embed_dim, export=export)
-        self.pointwise_conv1 = nn.Conv1d(embed_dim, 2 * channels, kernel_size=1, stride=1, padding=0, bias=bias)
-        self.glu = nn.GLU(dim=1)
-        self.depthwise_conv = nn.Conv1d(channels, channels, depthwise_kernel_size, stride=1, padding=(depthwise_kernel_size - 1) // 2, groups=channels, bias=bias)
-        self.batch_norm = nn.BatchNorm1d(channels)
-        self.activation = get_activation_fn(activation_fn)(channels)
-        self.pointwise_conv2 = nn.Conv1d(channels, embed_dim, kernel_size=1, stride=1, padding=0, bias=bias)
-        self.dropout = nn.Dropout(dropout)
-
-    def forward(self, x):
-        return self.dropout(self.pointwise_conv2(self.activation(self.batch_norm(self.depthwise_conv(self.glu(self.pointwise_conv1(self.layer_norm(x).transpose(1, 2)))))))).transpose(1, 2)
-
-def rotate_half(x):
-    x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]
-    return torch.cat((-x2, x1), dim=x1.ndim - 1)
-
-def apply_rotary_pos_emb(q, k, cos, sin, offset: int = 0):
-    cos, sin = (cos[offset : q.shape[0] + offset, ...], sin[offset : q.shape[0] + offset, ...])
-    return (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin)
-
-class RotaryPositionalEmbedding(nn.Module):
-    def __init__(self, dim, base=10000, precision=torch.half):
-        super().__init__()
-        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
-        self.register_buffer("inv_freq", inv_freq)
-        self.seq_len_cached = 0
-        self.cos_cached = torch.empty(self.seq_len_cached, 1, 1, dim)
-        self.sin_cached = torch.empty(self.seq_len_cached, 1, 1, dim)
-        self.precision = precision
-
-    def forward(self, x, seq_len = 0):
-        if seq_len > self.seq_len_cached:
-            self.seq_len_cached = seq_len
-            freqs = torch.einsum("i,j->ij", torch.arange(seq_len, device=x.device).type_as(self.inv_freq), self.inv_freq)
-            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
-            self.cos_cached = emb.cos().view(emb.size(0), 1, 1, emb.size(1))
-            self.sin_cached = emb.sin().view(emb.size(0), 1, 1, emb.size(1))
-        return self.cos_cached, self.sin_cached
-
-class ESPNETMultiHeadedAttention(nn.Module):
-    def __init__(self, n_feat, n_head, dropout):
-        super(ESPNETMultiHeadedAttention, self).__init__()
-        assert n_feat % n_head == 0
-        self.d_k = n_feat // n_head
-        self.h = n_head
-        self.linear_q = nn.Linear(n_feat, n_feat)
-        self.linear_k = nn.Linear(n_feat, n_feat)
-        self.linear_v = nn.Linear(n_feat, n_feat)
-        self.linear_out = nn.Linear(n_feat, n_feat)
-        self.attn = None
-        self.dropout = nn.Dropout(p=dropout)
-
-    def forward_qkv(self, query, key, value, **kwargs):
-        n_batch = query.size(0)
-        return self.linear_q(query).view(n_batch, -1, self.h, self.d_k).transpose(1, 2), self.linear_k(key).view(n_batch, -1, self.h, self.d_k).transpose(1, 2), self.linear_v(value).view(n_batch, -1, self.h, self.d_k).transpose(1, 2)
-
-    def forward_attention(self, value, scores, mask):
-        n_batch = value.size(0)
-
-        if mask is not None:
-            scores = scores.masked_fill(mask.unsqueeze(1).unsqueeze(2).to(bool), float("-inf"))
-            self.attn = torch.softmax(scores, dim=-1)
-        else: self.attn = torch.softmax(scores, dim=-1)
-
-        return self.linear_out((torch.matmul(self.dropout(self.attn), value).transpose(1, 2).contiguous().view(n_batch, -1, self.h * self.d_k))) 
-
-    def forward(self, query, key, value, key_padding_mask=None, **kwargs):
-        q, k, v = self.forward_qkv(query.transpose(0, 1), key.transpose(0, 1), value.transpose(0, 1))
-        return self.forward_attention(v, torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k), key_padding_mask).transpose(0, 1), None
-
-class RelPositionMultiHeadedAttention(ESPNETMultiHeadedAttention):
-    def __init__(self, n_feat, n_head, dropout, zero_triu=False):
-        super().__init__(n_feat, n_head, dropout)
-        self.zero_triu = zero_triu
-        self.linear_pos = nn.Linear(n_feat, n_feat, bias=False)
-        self.pos_bias_u = nn.Parameter(torch.zeros(self.h, self.d_k))
-        self.pos_bias_v = nn.Parameter(torch.zeros(self.h, self.d_k))
-        nn.init.xavier_uniform_(self.pos_bias_u)
-        nn.init.xavier_uniform_(self.pos_bias_v)
-
-    def rel_shift(self, x):
-        x = torch.cat([torch.zeros((*x.size()[:3], 1), device=x.device, dtype=x.dtype), x], dim=-1).view(*x.size()[:2], x.size(3) + 1, x.size(2))[:, :, 1:].view_as(x)[:, :, :, : x.size(-1) // 2 + 1]
-        if self.zero_triu: x = x * torch.tril(torch.ones((x.size(2), x.size(3)), device=x.device), x.size(3) - x.size(2))[None, None, :, :]
-        return x
-
-    def forward(self, query, key, value, pos_emb, key_padding_mask=None, **kwargs):
-        pos_emb = pos_emb.transpose(0, 1)
-        q, k, v = self.forward_qkv(query.transpose(0, 1), key.transpose(0, 1), value.transpose(0, 1))
-        q = q.transpose(1, 2)
-
-        return self.forward_attention(v, (torch.matmul((q + self.pos_bias_u).transpose(1, 2), k.transpose(-2, -1)) + self.rel_shift(torch.matmul((q + self.pos_bias_v).transpose(1, 2), self.linear_pos(pos_emb).view(pos_emb.size(0), -1, self.h, self.d_k).transpose(1, 2).transpose(-2, -1)))) / math.sqrt(self.d_k), key_padding_mask).transpose(0, 1), None
-
-class RotaryPositionMultiHeadedAttention(ESPNETMultiHeadedAttention):
-    def __init__(self, n_feat, n_head, dropout, precision, rotary_emd_base=10000):
-        super().__init__(n_feat, n_head, dropout)
-        precision = torch.float
-        self.rotary_ndims = self.d_k
-        if precision == "fp16": precision = torch.half
-        self.rotary_emb = RotaryPositionalEmbedding(self.rotary_ndims, base=rotary_emd_base, precision=precision)
-
-    def forward(self, query, key, value, key_padding_mask=None, **kwargs):
-        T, B, C = value.size()
-        query = query.view(T, B, self.h, self.d_k)
-        key = key.view(T, B, self.h, self.d_k)
-        value = value.view(T, B, self.h, self.d_k)
-
-        cos, sin = self.rotary_emb(value, seq_len=T)
-        query, key = apply_rotary_pos_emb(query, key, cos, sin, offset=0)
-
-        query = query.view(T, B, self.h * self.d_k)
-        key = key.view(T, B, self.h * self.d_k)
-        value = value.view(T, B, self.h * self.d_k)
-
-        q, k, v = self.forward_qkv(query.transpose(0, 1), key.transpose(0, 1), value.transpose(0, 1))
-        return self.forward_attention(v, torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k), key_padding_mask).transpose(0, 1), None
-
-class ConformerEncoderLayer(nn.Module):
-    def __init__(self, embed_dim, ffn_embed_dim, attention_heads, dropout, use_fp16, depthwise_conv_kernel_size=31, activation_fn="swish", attn_type=None, pos_enc_type="abs"):
-        self.pos_enc_type = pos_enc_type
-        super(ConformerEncoderLayer, self).__init__()
-        self.ffn1 = FeedForwardModule(embed_dim, ffn_embed_dim, dropout, dropout)
-        self.self_attn_layer_norm = LayerNorm(embed_dim, export=False)
-        self.self_attn_dropout = nn.Dropout(dropout)
-
-        if attn_type == "espnet":
-            if self.pos_enc_type == "rel_pos": self.self_attn = RelPositionMultiHeadedAttention(embed_dim, attention_heads, dropout=dropout)
-            elif self.pos_enc_type == "rope": self.self_attn = RotaryPositionMultiHeadedAttention(embed_dim, attention_heads, dropout=dropout, precision=use_fp16)
-            elif self.pos_enc_type == "abs": self.self_attn = ESPNETMultiHeadedAttention(embed_dim, attention_heads, dropout=dropout)
-            else: raise Exception
-        else: self.self_attn = MultiheadAttention(embed_dim, attention_heads, dropout=dropout)
-
-        self.conv_module = ConvolutionModule(embed_dim=embed_dim, channels=embed_dim, depthwise_kernel_size=depthwise_conv_kernel_size, dropout=dropout, activation_fn=activation_fn)
-        self.ffn2 = FeedForwardModule(embed_dim, ffn_embed_dim, dropout, dropout, activation_fn=activation_fn)
-        self.final_layer_norm = LayerNorm(embed_dim, export=False)
-
-    def forward(self, x, encoder_padding_mask, position_emb = None):
-        residual = x
-        x = self.ffn1(x) * 0.5 + residual
-        residual = x
-        x = self.self_attn_layer_norm(x)
-
-        if self.pos_enc_type == "rel_pos": x, attn = self.self_attn(query=x, key=x, value=x, key_padding_mask=encoder_padding_mask, pos_emb=position_emb, need_weights=False)
-        else: x, attn = self.self_attn(query=x, key=x, value=x, key_padding_mask=encoder_padding_mask, need_weights=False)
-
-        x = self.self_attn_dropout(x)
-        x = x + residual
-        residual = x
-        x = residual + self.conv_module(x.transpose(0, 1)).transpose(0, 1)
-        residual = x
-        x = self.ffn2(x)
-        layer_result = x
-        x = self.final_layer_norm(x * 0.5 + residual)
-
-        return x, (attn, layer_result)
-
-class ConformerWav2Vec2EncoderLayer(ConformerEncoderLayer):
-    def forward(self, x, self_attn_mask=None, self_attn_padding_mask=None, need_weights=False, att_args=None, position_emb=None):
-        return super().forward(x, self_attn_padding_mask, position_emb)
-
-class TransformerSentenceEncoderWithAdapterLayer(TransformerSentenceEncoderLayer):
-    def __init__(self, embedding_dim = 768, ffn_embedding_dim = 3072, num_attention_heads = 8, dropout = 0.1, attention_dropout = 0.1, activation_dropout = 0.1, activation_fn = "relu", layer_norm_first = False, adapter_num=201, adapter_dim=64, adapter_act_fn="relu"):
-        super().__init__(embedding_dim=embedding_dim, ffn_embedding_dim=ffn_embedding_dim, num_attention_heads=num_attention_heads, dropout=dropout, attention_dropout=attention_dropout, activation_dropout=activation_dropout, activation_fn=activation_fn, layer_norm_first=layer_norm_first)
-        self.adapter_num = adapter_num
-        self.adapter_dim = adapter_dim
-        self.adapter_layer = AdapterFast(adapter_num, self.embedding_dim, self.adapter_dim, adapter_act_fn)
-
-    def forward(self, x, self_attn_mask=None, self_attn_padding_mask=None, need_weights=False, att_args=None, corpus_key=None):
-
-        x, (attn, layer_result) = super().forward(x=x, self_attn_mask=self_attn_mask, self_attn_padding_mask=self_attn_padding_mask, need_weights=need_weights, att_args=att_args)
-        assert corpus_key is not None
-        assert len(set(corpus_key)) == 1
-
-        return x + self.adapter_layer(x, corpus_key[0]), (attn, layer_result)
-
-class TransposeLast(nn.Module):
-    def __init__(self, deconstruct_idx=None, tranpose_dim=-2):
-        super().__init__()
-        self.deconstruct_idx = deconstruct_idx
-        self.tranpose_dim = tranpose_dim
-
-    def forward(self, x):
-        if self.deconstruct_idx is not None: x = x[self.deconstruct_idx]
-        return x.transpose(self.tranpose_dim, -1)
-
-class TransformerEncoder(nn.Module):
-    def build_encoder_layer(self, args, **kwargs):
-        if args.layer_type == "transformer": layer = TransformerSentenceEncoderLayer(embedding_dim=self.embedding_dim, ffn_embedding_dim=args.encoder_ffn_embed_dim, num_attention_heads=args.encoder_attention_heads, dropout=self.dropout, attention_dropout=args.attention_dropout, activation_dropout=args.activation_dropout, activation_fn=args.activation_fn, layer_norm_first=args.layer_norm_first)
-        elif args.layer_type == "conformer": layer = ConformerWav2Vec2EncoderLayer(embed_dim=self.embedding_dim, ffn_embed_dim=args.encoder_ffn_embed_dim, attention_heads=args.encoder_attention_heads, dropout=args.dropout, depthwise_conv_kernel_size=args.depthwise_conv_kernel_size, activation_fn="swish", attn_type=args.attn_type, use_fp16=args.fp16, pos_enc_type="abs")
-        elif args.layer_type == "trf_adp":
-            use_adp = False
-            if args.adp_trf_idx == "all": use_adp = True
-            else: 
-                if kwargs.get("layer_idx", None) in list(range(*[int(g) for g in args.adp_trf_idx.split(":")])): use_adp = True
-
-            layer = TransformerSentenceEncoderWithAdapterLayer(embedding_dim=self.embedding_dim, ffn_embedding_dim=args.encoder_ffn_embed_dim, num_attention_heads=args.encoder_attention_heads, dropout=self.dropout, attention_dropout=args.attention_dropout, activation_dropout=args.activation_dropout, activation_fn=args.activation_fn, layer_norm_first=args.layer_norm_first, adapter_num=args.adp_num, adapter_dim=args.adp_dim, adapter_act_fn=args.adp_act_fn) if use_adp else TransformerSentenceEncoderLayer(embedding_dim=self.embedding_dim, ffn_embedding_dim=args.encoder_ffn_embed_dim, num_attention_heads=args.encoder_attention_heads, dropout=self.dropout, attention_dropout=args.attention_dropout, activation_dropout=args.activation_dropout, activation_fn=args.activation_fn, layer_norm_first=args.layer_norm_first,)
-
-        return layer
-
-    def __init__(self, args):
-        super().__init__()
-        self.dropout = args.dropout
-        self.embedding_dim = args.encoder_embed_dim
-        self.required_seq_len_multiple = args.required_seq_len_multiple
-        pos_conv_depth = getattr(args, "pos_conv_depth", 1)
-
-        if pos_conv_depth > 1:
-            num_layers = args.pos_conv_depth
-            k = max(3, args.conv_pos // num_layers)
-
-            def make_conv_block(e, k, g, l):
-                return nn.Sequential(*[nn.Sequential(nn.Conv1d(e, e, kernel_size=k, padding=k // 2, groups=g), SamePad(k), TransposeLast(), LayerNorm(e, elementwise_affine=False), TransposeLast(), nn.GELU()) for _ in range(l)])
-
-            self.pos_conv = make_conv_block(self.embedding_dim, k, args.conv_pos_groups, num_layers)
-        else: self.pos_conv = make_conv_pos(self.embedding_dim, args.conv_pos, args.conv_pos_groups)
-
-        self.layers = nn.ModuleList([self.build_encoder_layer(args, layer_idx=ii) for ii in range(args.encoder_layers)])
-        self.layer_norm_first = args.layer_norm_first
-        self.layer_norm = LayerNorm(self.embedding_dim)
-        self.layerdrop = args.encoder_layerdrop
-        self.apply(init_bert_params)
-
-    def forward(self, x, padding_mask=None, layer=None, corpus_key=None):
-        x, layer_results = self.extract_features(x, padding_mask, layer, corpus_key=corpus_key)
-
-        if self.layer_norm_first and layer is None: x = self.layer_norm(x)
-        return x, layer_results
-
-    def extract_features(self, x, padding_mask=None, tgt_layer=None, min_layer=0, corpus_key=None):
-        if padding_mask is not None: x = index_put(x, padding_mask, 0)
-        x = x + self.pos_conv(x.transpose(1, 2)).transpose(1, 2)
-
-        if not self.layer_norm_first: x = self.layer_norm(x)
-        x, pad_length = pad_to_multiple(x, self.required_seq_len_multiple, dim=-2, value=0)
-
-        if pad_length > 0 and padding_mask is None:
-            padding_mask = x.new_zeros((x.size(0), x.size(1)), dtype=torch.bool)
-            padding_mask[:, -pad_length:] = True
-        else: padding_mask, _ = pad_to_multiple(padding_mask, self.required_seq_len_multiple, dim=-1, value=True)
-
-        x = F.dropout(x, p=self.dropout, training=self.training).transpose(0, 1)
-        layer_results = []
-        r = None
-
-        for i, layer in enumerate(self.layers):
-            dropout_probability = np.random.random() if self.layerdrop > 0 else 1
-            if not self.training or (dropout_probability > self.layerdrop):
-                layer_check = layer
-
-                if (corpus_key is None) or (not isinstance(layer_check, (TransformerSentenceEncoderWithAdapterLayer))): x, (z, lr) = layer(x, self_attn_padding_mask=padding_mask, need_weights=False)
-                else: x, (z, lr) = layer(x, self_attn_padding_mask=padding_mask, need_weights=False, corpus_key=corpus_key)
-
-                if i >= min_layer: layer_results.append((x, z, lr))
-            if i == tgt_layer:
-                r = x
-                break
-
-        if r is not None: x = r
-        x = x.transpose(0, 1)
-
-        if pad_length > 0:
-            x = x[:, :-pad_length]
-            def undo_pad(a, b, c):
-                return (a[:-pad_length], b[:-pad_length] if b is not None else b, c[:-pad_length])
-
-            layer_results = [undo_pad(*u) for u in layer_results]
-
-        return x, layer_results
-
-    def max_positions(self):
-        return self.args.max_positions
-
-    def upgrade_state_dict_named(self, state_dict, name):
-        return state_dict
-
-class Fp32GroupNorm(nn.GroupNorm):
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-    def forward(self, input):
-        output = F.group_norm(input.float(), self.num_groups, self.weight.float() if self.weight is not None else None, self.bias.float() if self.bias is not None else None, self.eps)
-        return output.type_as(input)
-
-class Fp32LayerNorm(nn.LayerNorm):
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-    def forward(self, input):
-        output = F.layer_norm(input.float(), self.normalized_shape, self.weight.float() if self.weight is not None else None, self.bias.float() if self.bias is not None else None, self.eps)
-        return output.type_as(input)
-
-class ConvFeatureExtractionModel(nn.Module):
-    def __init__(self, conv_layers, dropout = 0.0, mode = "default", conv_bias = False):
-        super().__init__()
-        assert mode in {"default", "layer_norm"}
-
-        def block(n_in, n_out, k, stride, is_layer_norm=False, is_group_norm=False, conv_bias=False):
-            def make_conv():
-                conv = nn.Conv1d(n_in, n_out, k, stride=stride, bias=conv_bias)
-                nn.init.kaiming_normal_(conv.weight)
-                return conv
-
-            assert (is_layer_norm and is_group_norm) == False
-
-            if is_layer_norm: return nn.Sequential(make_conv(), nn.Dropout(p=dropout), nn.Sequential(TransposeLast(), Fp32LayerNorm(dim, elementwise_affine=True), TransposeLast()), nn.GELU())
-            elif is_group_norm: return nn.Sequential(make_conv(), nn.Dropout(p=dropout), Fp32GroupNorm(dim, dim, affine=True), nn.GELU())
-            else: return nn.Sequential(make_conv(), nn.Dropout(p=dropout), nn.GELU())
-
-        in_d = 1
-        self.conv_layers = nn.ModuleList()
-        for i, cl in enumerate(conv_layers):
-            assert len(cl) == 3
-            (dim, k, stride) = cl
-            self.conv_layers.append(block(in_d, dim, k, stride, is_layer_norm=mode == "layer_norm", is_group_norm=mode == "default" and i == 0, conv_bias=conv_bias))
-            in_d = dim
-
-    def forward(self, x):
-        x = x.unsqueeze(1)
-        for conv in self.conv_layers:
-            x = conv(x)
-
-        return x
-
-class GradMultiply(torch.autograd.Function):
-    @staticmethod
-    def forward(ctx, x, scale):
-        ctx.scale = scale
-        res = x.new(x)
-        return res
-
-    @staticmethod
-    def backward(ctx, grad):
-        return grad * ctx.scale, None
-
-class BaseFairseqModel(nn.Module):
-    def __init__(self):
-        super().__init__()
-        self._is_generation_fast = False
-
-    def get_targets(self, sample, net_output):
-        return sample["target"]
-
-    def extract_features(self, *args, **kwargs):
-        return self(*args, **kwargs)
-
-    def load_state_dict(self, state_dict, strict=True, model_cfg = None, args = None):
-        self.upgrade_state_dict(state_dict)
-        new_state_dict = prune_state_dict(state_dict, model_cfg)
-        return super().load_state_dict(new_state_dict, strict)
-
-    def upgrade_state_dict(self, state_dict):
-        self.upgrade_state_dict_named(state_dict, "")
-
-    def upgrade_state_dict_named(self, state_dict, name):
-        assert state_dict is not None
-
-        def do_upgrade(m, prefix):
-            if len(prefix) > 0: prefix += "."
-
-            for n, c in m.named_children():
-                name = prefix + n
-                if hasattr(c, "upgrade_state_dict_named"): c.upgrade_state_dict_named(state_dict, name)
-                elif hasattr(c, "upgrade_state_dict"): c.upgrade_state_dict(state_dict)
-                do_upgrade(c, name)
-
-        do_upgrade(self, name)
-
-    def make_generation_fast_(self, **kwargs):
-        if self._is_generation_fast: return
-        self._is_generation_fast = True
-
-        def apply_remove_weight_norm(module):
-            try:
-                nn.utils.remove_weight_norm(module)
-            except (AttributeError, ValueError):
-                return
-
-        self.apply(apply_remove_weight_norm)
-
-        def apply_make_generation_fast_(module, prefix):
-            if len(prefix) > 0: prefix += "."
-
-            base_func = BaseFairseqModel.make_generation_fast_
-            for n, m in module.named_modules():
-                if (m != self and hasattr(m, "make_generation_fast_") and m.make_generation_fast_.__func__ is not base_func): m.make_generation_fast_(name=prefix + n, **kwargs)
-
-        apply_make_generation_fast_(self, "")
-        self.eval()
-
-class HubertConfig:
-    def __init__(self, _name, label_rate, encoder_layers_1, logit_temp_ctr, num_negatives, cross_sample_negatives, ctr_layers, extractor_mode = "default", encoder_layers = 12, encoder_embed_dim = 768, encoder_ffn_embed_dim = 3072, encoder_attention_heads = 12, activation_fn = "gelu", layer_type = "transformer", dropout = 0.1, attention_dropout = 0.1, activation_dropout = 0.0, encoder_layerdrop = 0.0, dropout_input = 0.0, dropout_features = 0.0, final_dim = 0, untie_final_proj = False, layer_norm_first = False, conv_feature_layers = "[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2", conv_bias = False, logit_temp = 0.1, target_glu = False, feature_grad_mult = 1.0, mask_length = 10, mask_prob = 0.65, mask_selection = "static", mask_other = 0.0, no_mask_overlap = False, mask_min_space = 1, mask_channel_length = 10, mask_channel_prob = 0.0, mask_channel_selection = "static", mask_channel_other = 0.0, no_mask_channel_overlap = False, mask_channel_min_space = 1, conv_pos = 128, conv_pos_groups = 16, conv_pos_batch_norm = False, latent_temp = (2, 0.5, 0.999995), skip_masked = False, skip_nomask = False, checkpoint_activations = False, required_seq_len_multiple = 2, depthwise_conv_kernel_size = 31, attn_type = "", pos_enc_type = "abs", fp16 = False):
-        self._name = _name
-        self.label_rate = label_rate
-        self.encoder_layers_1 = encoder_layers_1
-        self.logit_temp_ctr = logit_temp_ctr
-        self.num_negatives = num_negatives
-        self.cross_sample_negatives = cross_sample_negatives
-        self.ctr_layers = ctr_layers
-        self.extractor_mode = extractor_mode
-        self.encoder_layers = encoder_layers
-        self.encoder_embed_dim = encoder_embed_dim
-        self.encoder_ffn_embed_dim = encoder_ffn_embed_dim
-        self.encoder_attention_heads = encoder_attention_heads
-        self.activation_fn = activation_fn
-        self.layer_type = layer_type
-        self.dropout = dropout
-        self.attention_dropout = attention_dropout
-        self.activation_dropout = activation_dropout
-        self.encoder_layerdrop = encoder_layerdrop
-        self.dropout_input = encoder_layerdrop
-        self.dropout_features = dropout_features
-        self.final_dim = final_dim
-        self.untie_final_proj = untie_final_proj
-        self.layer_norm_first = layer_norm_first
-        self.conv_feature_layers = conv_feature_layers
-        self.conv_bias = conv_bias
-        self.logit_temp = logit_temp
-        self.target_glu = target_glu
-        self.feature_grad_mult = feature_grad_mult
-        self.mask_length = mask_length
-        self.mask_prob = mask_prob
-        self.mask_selection = mask_selection
-        self.mask_other = mask_other
-        self.no_mask_overlap = no_mask_overlap
-        self.mask_min_space = mask_min_space
-        self.mask_channel_length = mask_channel_length
-        self.mask_channel_prob = mask_channel_prob
-        self.mask_channel_selection = mask_channel_selection
-        self.mask_channel_other = mask_channel_other
-        self.no_mask_channel_overlap = no_mask_channel_overlap
-        self.mask_channel_min_space = mask_channel_min_space
-        self.conv_pos = conv_pos
-        self.conv_pos_groups = conv_pos_groups
-        self.conv_pos_batch_norm = conv_pos_batch_norm
-        self.latent_temp = latent_temp
-        self.skip_masked = skip_masked
-        self.skip_nomask = skip_nomask
-        self.checkpoint_activations = checkpoint_activations
-        self.required_seq_len_multiple = required_seq_len_multiple
-        self.depthwise_conv_kernel_size = depthwise_conv_kernel_size
-        self.attn_type = attn_type
-        self.pos_enc_type = pos_enc_type
-        self.fp16 = fp16
-
-class Model_Config(dict):
-    def __getattr__(*args):
-        val = dict.get(*args)
-        return Model_Config(val) if type(val) is dict else val
-
-    __setattr__ = dict.__setitem__
-    __delattr__ = dict.__delitem__  
-
-class HubertModel(BaseFairseqModel):
-    def __init__(self, cfg):
-        super().__init__()
-        feature_enc_layers = eval(cfg.conv_feature_layers)
-        self.embed = feature_enc_layers[-1][0]
-        self.feature_extractor = ConvFeatureExtractionModel(conv_layers=feature_enc_layers, dropout=0.0, mode=cfg.extractor_mode, conv_bias=cfg.conv_bias)
-        feature_ds_rate = np.prod([s for _, _, s in feature_enc_layers])
-        self.feat2tar_ratio = cfg.label_rate * feature_ds_rate / 16000
-        self.post_extract_proj = (nn.Linear(self.embed, cfg.encoder_embed_dim) if self.embed != cfg.encoder_embed_dim else None)
-        self.mask_prob = cfg.mask_prob
-        self.mask_selection = cfg.mask_selection
-        self.mask_other = cfg.mask_other
-        self.mask_length = cfg.mask_length
-        self.no_mask_overlap = cfg.no_mask_overlap
-        self.mask_min_space = cfg.mask_min_space
-        self.mask_channel_prob = cfg.mask_channel_prob
-        self.mask_channel_selection = cfg.mask_channel_selection
-        self.mask_channel_other = cfg.mask_channel_other
-        self.mask_channel_length = cfg.mask_channel_length
-        self.no_mask_channel_overlap = cfg.no_mask_channel_overlap
-        self.mask_channel_min_space = cfg.mask_channel_min_space
-        self.dropout_input = nn.Dropout(cfg.dropout_input)
-        self.dropout_features = nn.Dropout(cfg.dropout_features)
-        self.feature_grad_mult = cfg.feature_grad_mult
-        self.logit_temp = cfg.logit_temp
-        self.skip_masked = cfg.skip_masked
-        self.skip_nomask = cfg.skip_nomask
-        final_dim = cfg.final_dim if cfg.final_dim > 0 else cfg.encoder_embed_dim
-        self.mask_emb = nn.Parameter(torch.FloatTensor(cfg.encoder_embed_dim).uniform_())
-        self.encoder = TransformerEncoder(cfg)
-        self.layer_norm = LayerNorm(self.embed)
-        self.target_glu = None
-        if cfg.target_glu: self.target_glu = nn.Sequential(nn.Linear(final_dim, final_dim * 2), nn.GLU())
-        self.untie_final_proj = cfg.untie_final_proj
-        self.final_proj = nn.Linear(cfg.encoder_embed_dim, final_dim)
-        self.num_classes = [504]
-        self.label_embs_concat = nn.Parameter(torch.FloatTensor(sum(self.num_classes), final_dim))
-        nn.init.uniform_(self.label_embs_concat)
-
-    def upgrade_state_dict_named(self, state_dict, name):
-        super().upgrade_state_dict_named(state_dict, name)
-        return state_dict
-
-    def apply_mask(self, x, padding_mask, target_list):
-        B, T, C = x.shape
-        if self.mask_prob > 0:
-            mask_indices = torch.from_numpy(compute_mask_indices((B, T), padding_mask, self.mask_prob, self.mask_length, self.mask_selection, self.mask_other, min_masks=2, no_overlap=self.no_mask_overlap, min_space=self.mask_min_space)).to(x.device)
-            x[mask_indices] = self.mask_emb
-        else: mask_indices = None
-
-        if self.mask_channel_prob > 0: x[(torch.from_numpy(compute_mask_indices((B, C), None, self.mask_channel_prob, self.mask_channel_length, self.mask_channel_selection, self.mask_channel_other, no_overlap=self.no_mask_channel_overlap, min_space=self.mask_channel_min_space)).to(x.device).unsqueeze(1).expand(-1, T, -1))] = 0
-        return x, mask_indices
-
-    def compute_nce(self, x, pos, negs):
-        neg_is_pos = (pos == negs).all(-1)
-        logits = torch.cosine_similarity(x.float(), torch.cat([pos.unsqueeze(0), negs], dim=0).float(), dim=-1).type_as(x)
-        logits /= self.logit_temp
-
-        if neg_is_pos.any(): logits[1:][neg_is_pos] = float("-inf")
-        return logits.transpose(0, 1)
-
-    def forward_features(self, source):
-        if self.feature_grad_mult > 0:
-            features = self.feature_extractor(source)
-            if self.feature_grad_mult != 1.0: features = GradMultiply.apply(features, self.feature_grad_mult)
-        else:
-            with torch.no_grad():
-                features = self.feature_extractor(source)
-        return features
-
-    def forward_targets(self, features, target_list):
-        feat_tsz = features.size(2)
-        targ_tsz = min([t.size(1) for t in target_list])
-
-        if self.feat2tar_ratio * feat_tsz > targ_tsz:
-            feat_tsz = int(targ_tsz / self.feat2tar_ratio)
-            features = features[..., :feat_tsz]
-
-        return features, [t[:, (torch.arange(feat_tsz).float() * self.feat2tar_ratio).long()] for t in target_list]
-
-    def forward_padding_mask(self, features, padding_mask):
-        extra = padding_mask.size(1) % features.size(1)
-        if extra > 0: padding_mask = padding_mask[:, :-extra]
-
-        return padding_mask.view(padding_mask.size(0), features.size(1), -1).all(-1)
-
-    def forward(self, source, target_list = None, padding_mask = None, mask = True, features_only = False, output_layer = None):
-        features = self.forward_features(source)
-        if target_list is not None: features, target_list = self.forward_targets(features, target_list)
-
-        features_pen = features.float().pow(2).mean()
-
-        features = self.layer_norm(features.transpose(1, 2))
-        unmasked_features = features.clone()
-
-        if padding_mask is not None: padding_mask = self.forward_padding_mask(features, padding_mask)
-        if self.post_extract_proj is not None: features = self.post_extract_proj(features)
-
-        features = self.dropout_input(features)
-        unmasked_features = self.dropout_features(unmasked_features)
-
-        if mask: x, mask_indices = self.apply_mask(features, padding_mask, target_list)
-        else: x, mask_indices = features, None
-
-        x, _ = self.encoder(x, padding_mask=padding_mask, layer=None if output_layer is None else output_layer - 1)
-        if features_only: return {"x": x, "padding_mask": padding_mask, "features": features}
-
-        def compute_pred(proj_x, target, label_embs):
-            y = torch.index_select(label_embs, 0, target.long())
-            negs = label_embs.unsqueeze(1).expand(-1, proj_x.size(0), -1)
-
-            if self.target_glu:
-                y = self.target_glu(y)
-                negs = self.target_glu(negs)
-
-            return self.compute_nce(proj_x, y, negs)
-
-        label_embs_list = self.label_embs_concat.split(self.num_classes, 0)
-
-        if not self.skip_masked:
-            masked_indices = torch.logical_and(~padding_mask, mask_indices)
-            proj_x_m = self.final_proj(x[masked_indices])
-            logit_m_list = [compute_pred(proj_x_m, t[masked_indices], label_embs_list[i]) for i, (proj_x_m, t) in enumerate(zip(proj_x_m.chunk(len(target_list), dim=-1) if self.untie_final_proj else [proj_x_m for _ in range(len(target_list))], target_list))]
-        else: logit_m_list = [None for _ in target_list]
-
-        if not self.skip_nomask:
-            nomask_indices = torch.logical_and(~padding_mask, ~mask_indices)
-            proj_x_u = self.final_proj(x[nomask_indices])
-            logit_u_list = [compute_pred(proj_x_u, t[nomask_indices], label_embs_list[i]) for i, (proj_x_u, t) in enumerate(zip(proj_x_u.chunk(len(target_list), dim=-1) if self.untie_final_proj else [proj_x_u for _ in range(len(target_list))], target_list))]
-        else: logit_u_list = [None for _ in target_list]
-
-        return {"logit_m_list": logit_m_list, "logit_u_list": logit_u_list, "padding_mask": padding_mask, "features_pen": features_pen}
-
-    def extract_features(self, source, padding_mask = None, mask = False, ret_conv = False, output_layer = None):
-        res = self.forward(source, padding_mask=padding_mask, mask=mask, features_only=True, output_layer=output_layer)
-        return res["features"] if ret_conv else res["x"], res["padding_mask"]
-
-    def get_logits(self, net_output, is_masked=True):
-        return [x.float() for x in (net_output["logit_m_list"] if is_masked else net_output["logit_u_list"]) if x is not None]
-
-    def get_targets(self, net_output, is_masked=True):
-        return [x.new_zeros(x.size(0), dtype=torch.long) for x in self.get_logits(net_output, is_masked)]
-
-    def get_extra_losses(self, net_output):
-        extra_losses, names = [], []
-
-        if "features_pen" in net_output:
-            extra_losses.append(net_output["features_pen"])
-            names.append("features_pen")
-
-        return extra_losses, names
-
-    def remove_pretraining_modules(self):
-        self.target_glu = None
-        self.final_proj = None

main/library/architectures/mdx_separator.py

@@ -1,320 +0,0 @@
-import os
-import sys
-import onnx
-import torch
-import platform
-import onnx2torch
-
-import numpy as np
-import onnxruntime as ort
-
-from tqdm import tqdm
-
-sys.path.append(os.getcwd())
-
-from main.configs.config import Config
-from main.library.uvr5_separator import spec_utils
-from main.library.uvr5_separator.common_separator import CommonSeparator
-
-translations = Config().translations
-
-class MDXSeparator(CommonSeparator):
-    def __init__(self, common_config, arch_config):
-        super().__init__(config=common_config)
-        self.segment_size = arch_config.get("segment_size")
-        self.overlap = arch_config.get("overlap")
-        self.batch_size = arch_config.get("batch_size", 1)
-        self.hop_length = arch_config.get("hop_length")
-        self.enable_denoise = arch_config.get("enable_denoise")
-        self.logger.debug(translations["mdx_info"].format(batch_size=self.batch_size, segment_size=self.segment_size))
-        self.logger.debug(translations["mdx_info_2"].format(overlap=self.overlap, hop_length=self.hop_length, enable_denoise=self.enable_denoise))
-        self.compensate = self.model_data["compensate"]
-        self.dim_f = self.model_data["mdx_dim_f_set"]
-        self.dim_t = 2 ** self.model_data["mdx_dim_t_set"]
-        self.n_fft = self.model_data["mdx_n_fft_scale_set"]
-        self.config_yaml = self.model_data.get("config_yaml", None)
-        self.logger.debug(f"{translations['mdx_info_3']}: compensate = {self.compensate}, dim_f = {self.dim_f}, dim_t = {self.dim_t}, n_fft = {self.n_fft}")
-        self.logger.debug(f"{translations['mdx_info_3']}: config_yaml = {self.config_yaml}")
-        self.load_model()
-        self.n_bins = 0
-        self.trim = 0
-        self.chunk_size = 0
-        self.gen_size = 0
-        self.stft = None
-        self.primary_source = None
-        self.secondary_source = None
-        self.audio_file_path = None
-        self.audio_file_base = None
-
-    def load_model(self):
-        self.logger.debug(translations["load_model_onnx"])
-
-        if self.segment_size == self.dim_t:
-            ort_session_options = ort.SessionOptions()
-            ort_session_options.log_severity_level = 3 if self.log_level > 10 else 0
-            ort_inference_session = ort.InferenceSession(self.model_path, providers=self.onnx_execution_provider, sess_options=ort_session_options)
-            self.model_run = lambda spek: ort_inference_session.run(None, {"input": spek.cpu().numpy()})[0]
-            self.logger.debug(translations["load_model_onnx_success"])
-        else:
-            self.model_run = onnx2torch.convert(onnx.load(self.model_path)) if platform.system() == 'Windows' else onnx2torch.convert(self.model_path)
-            self.model_run.to(self.torch_device).eval()
-            self.logger.debug(translations["onnx_to_pytorch"])
-
-    def separate(self, audio_file_path):
-        self.audio_file_path = audio_file_path
-        self.audio_file_base = os.path.splitext(os.path.basename(audio_file_path))[0]
-        self.logger.debug(translations["mix"].format(audio_file_path=self.audio_file_path))
-        mix = self.prepare_mix(self.audio_file_path)
-        self.logger.debug(translations["normalization_demix"])
-        mix = spec_utils.normalize(wave=mix, max_peak=self.normalization_threshold)
-        source = self.demix(mix)
-        self.logger.debug(translations["mix_success"])
-        output_files = []
-        self.logger.debug(translations["process_output_file"])
-
-        if not isinstance(self.primary_source, np.ndarray):
-            self.logger.debug(translations["primary_source"])
-            self.primary_source = spec_utils.normalize(wave=source, max_peak=self.normalization_threshold).T
-
-        if not isinstance(self.secondary_source, np.ndarray):
-            self.logger.debug(translations["secondary_source"])
-            raw_mix = self.demix(mix, is_match_mix=True)
-
-            if self.invert_using_spec:
-                self.logger.debug(translations["invert_using_spec"])
-                self.secondary_source = spec_utils.invert_stem(raw_mix, source)
-            else:
-                self.logger.debug(translations["invert_using_spec_2"])
-                self.secondary_source = mix.T - source.T
-
-        if not self.output_single_stem or self.output_single_stem.lower() == self.secondary_stem_name.lower():
-            self.secondary_stem_output_path = os.path.join(f"{self.audio_file_base}_({self.secondary_stem_name})_{self.model_name}.{self.output_format.lower()}")
-            self.logger.info(translations["save_secondary_stem_output_path"].format(stem_name=self.secondary_stem_name, stem_output_path=self.secondary_stem_output_path))
-            self.final_process(self.secondary_stem_output_path, self.secondary_source, self.secondary_stem_name)
-            output_files.append(self.secondary_stem_output_path)
-
-        if not self.output_single_stem or self.output_single_stem.lower() == self.primary_stem_name.lower():
-            self.primary_stem_output_path = os.path.join(f"{self.audio_file_base}_({self.primary_stem_name})_{self.model_name}.{self.output_format.lower()}")
-            if not isinstance(self.primary_source, np.ndarray): self.primary_source = source.T
-
-            self.logger.info(translations["save_secondary_stem_output_path"].format(stem_name=self.primary_stem_name, stem_output_path=self.primary_stem_output_path))
-            self.final_process(self.primary_stem_output_path, self.primary_source, self.primary_stem_name)
-            output_files.append(self.primary_stem_output_path)
-
-        return output_files
-
-    def initialize_model_settings(self):
-        self.logger.debug(translations["starting_model"])
-
-        self.n_bins = self.n_fft // 2 + 1
-        self.trim = self.n_fft // 2
-
-        self.chunk_size = self.hop_length * (self.segment_size - 1)
-        self.gen_size = self.chunk_size - 2 * self.trim
-
-        self.stft = STFT(self.logger, self.n_fft, self.hop_length, self.dim_f, self.torch_device)
-
-        self.logger.debug(f"{translations['input_info']}: n_fft = {self.n_fft} hop_length = {self.hop_length} dim_f = {self.dim_f}")
-        self.logger.debug(f"{translations['model_settings']}: n_bins = {self.n_bins}, Trim = {self.trim}, chunk_size = {self.chunk_size}, gen_size = {self.gen_size}")
-
-    def initialize_mix(self, mix, is_ckpt=False):
-        self.logger.debug(translations["initialize_mix"].format(is_ckpt=is_ckpt, shape=mix.shape))
-
-        if mix.shape[0] != 2:
-            error_message = translations["!=2"].format(shape=mix.shape[0])
-            self.logger.error(error_message)
-            raise ValueError(error_message)
-
-        if is_ckpt:
-            self.logger.debug(translations["process_check"])
-            pad = self.gen_size + self.trim - (mix.shape[-1] % self.gen_size)
-            self.logger.debug(f"{translations['cache']}: {pad}")
-
-            mixture = np.concatenate((np.zeros((2, self.trim), dtype="float32"), mix, np.zeros((2, pad), dtype="float32")), 1)
-
-            num_chunks = mixture.shape[-1] // self.gen_size
-            self.logger.debug(translations["shape"].format(shape=mixture.shape, num_chunks=num_chunks))
-
-            mix_waves = [mixture[:, i * self.gen_size : i * self.gen_size + self.chunk_size] for i in range(num_chunks)]
-        else:
-            self.logger.debug(translations["process_no_check"])
-            mix_waves = []
-            n_sample = mix.shape[1]
-
-            pad = self.gen_size - n_sample % self.gen_size
-            self.logger.debug(translations["n_sample_or_pad"].format(n_sample=n_sample, pad=pad))
-
-            mix_p = np.concatenate((np.zeros((2, self.trim)), mix, np.zeros((2, pad)), np.zeros((2, self.trim))), 1)
-            self.logger.debug(f"{translations['shape_2']}: {mix_p.shape}")
-
-            i = 0
-            while i < n_sample + pad:
-                mix_waves.append(np.array(mix_p[:, i : i + self.chunk_size]))
-
-                self.logger.debug(translations["process_part"].format(mix_waves=len(mix_waves), i=i, ii=i + self.chunk_size))
-                i += self.gen_size
-
-        mix_waves_tensor = torch.tensor(mix_waves, dtype=torch.float32).to(self.torch_device)
-        self.logger.debug(translations["mix_waves_to_tensor"].format(shape=mix_waves_tensor.shape))
-
-        return mix_waves_tensor, pad
-
-    def demix(self, mix, is_match_mix=False):
-        self.logger.debug(f"{translations['demix_is_match_mix']}: {is_match_mix}...")
-        self.initialize_model_settings()
-        self.logger.debug(f"{translations['mix_shape']}: {mix.shape}")
-        tar_waves_ = []
-
-        if is_match_mix:
-            chunk_size = self.hop_length * (self.segment_size - 1)
-            overlap = 0.02
-            self.logger.debug(translations["chunk_size_or_overlap"].format(chunk_size=chunk_size, overlap=overlap))
-        else:
-            chunk_size = self.chunk_size
-            overlap = self.overlap
-            self.logger.debug(translations["chunk_size_or_overlap_standard"].format(chunk_size=chunk_size, overlap=overlap))
-
-        gen_size = chunk_size - 2 * self.trim
-        self.logger.debug(f"{translations['calc_size']}: {gen_size}")
-
-        mixture = np.concatenate((np.zeros((2, self.trim), dtype="float32"), mix, np.zeros((2, gen_size + self.trim - ((mix.shape[-1]) % gen_size)), dtype="float32")), 1)
-        self.logger.debug(f"{translations['mix_cache']}: {mixture.shape}")
-
-        step = int((1 - overlap) * chunk_size)
-        self.logger.debug(translations["step_or_overlap"].format(step=step, overlap=overlap))
-
-        result = np.zeros((1, 2, mixture.shape[-1]), dtype=np.float32)
-        divider = np.zeros((1, 2, mixture.shape[-1]), dtype=np.float32)
-
-        total = 0
-        total_chunks = (mixture.shape[-1] + step - 1) // step
-        self.logger.debug(f"{translations['all_process_part']}: {total_chunks}")
-
-        for i in tqdm(range(0, mixture.shape[-1], step), ncols=100, unit="f"):
-            total += 1
-            start = i
-            end = min(i + chunk_size, mixture.shape[-1])
-            self.logger.debug(translations["process_part_2"].format(total=total, total_chunks=total_chunks, start=start, end=end))
-
-            chunk_size_actual = end - start
-            window = None
-
-            if overlap != 0:
-                window = np.hanning(chunk_size_actual)
-                window = np.tile(window[None, None, :], (1, 2, 1))
-                self.logger.debug(translations["window"])
-
-            mix_part_ = mixture[:, start:end]
-            
-            if end != i + chunk_size:
-                pad_size = (i + chunk_size) - end
-                mix_part_ = np.concatenate((mix_part_, np.zeros((2, pad_size), dtype="float32")), axis=-1)
-
-            mix_waves = torch.tensor([mix_part_], dtype=torch.float32).to(self.torch_device).split(self.batch_size)
-
-            total_batches = len(mix_waves)
-            self.logger.debug(f"{translations['mix_or_batch']}: {total_batches}")
-
-            with torch.no_grad():
-                batches_processed = 0
-                
-                for mix_wave in mix_waves:
-                    batches_processed += 1
-                    self.logger.debug(f"{translations['mix_wave']} {batches_processed}/{total_batches}")
-
-                    tar_waves = self.run_model(mix_wave, is_match_mix=is_match_mix)
-
-                    if window is not None:
-                        tar_waves[..., :chunk_size_actual] *= window
-                        divider[..., start:end] += window
-                    else: divider[..., start:end] += 1
-
-                    result[..., start:end] += tar_waves[..., : end - start]
-
-
-        self.logger.debug(translations["normalization_2"])
-        tar_waves = result / divider
-        tar_waves_.append(tar_waves)
-
-        tar_waves = np.concatenate(np.vstack(tar_waves_)[:, :, self.trim : -self.trim], axis=-1)[:, : mix.shape[-1]]
-
-        source = tar_waves[:, 0:None]
-        self.logger.debug(f"{translations['tar_waves']}: {tar_waves.shape}")
-
-        if not is_match_mix:
-            source *= self.compensate
-            self.logger.debug(translations["mix_match"])
-
-        self.logger.debug(translations["mix_success"])
-        return source
-
-    def run_model(self, mix, is_match_mix=False):
-        spek = self.stft(mix.to(self.torch_device))
-        self.logger.debug(translations["stft_2"].format(shape=spek.shape))
-
-        spek[:, :, :3, :] *= 0
-
-        if is_match_mix:
-            spec_pred = spek.cpu().numpy()
-            self.logger.debug(translations["is_match_mix"])
-        else:
-            if self.enable_denoise:
-                spec_pred_neg = self.model_run(-spek)  
-                spec_pred_pos = self.model_run(spek)
-                spec_pred = (spec_pred_neg * -0.5) + (spec_pred_pos * 0.5)
-                self.logger.debug(translations["enable_denoise"])
-            else:
-                spec_pred = self.model_run(spek)
-                self.logger.debug(translations["no_denoise"])
-
-        result = self.stft.inverse(torch.tensor(spec_pred).to(self.torch_device)).cpu().detach().numpy()
-        self.logger.debug(f"{translations['stft']}: {result.shape}")
-
-        return result
-
-class STFT:
-    def __init__(self, logger, n_fft, hop_length, dim_f, device):
-        self.logger = logger
-        self.n_fft = n_fft
-        self.hop_length = hop_length
-        self.dim_f = dim_f
-        self.device = device
-        self.hann_window = torch.hann_window(window_length=self.n_fft, periodic=True)
-
-    def __call__(self, input_tensor):
-        is_non_standard_device = not input_tensor.device.type in ["cuda", "cpu"]
-
-        if is_non_standard_device: input_tensor = input_tensor.cpu()
-
-        batch_dimensions = input_tensor.shape[:-2]
-        channel_dim, time_dim = input_tensor.shape[-2:]
-
-        permuted_stft_output = torch.stft(input_tensor.reshape([-1, time_dim]), n_fft=self.n_fft, hop_length=self.hop_length, window=self.hann_window.to(input_tensor.device), center=True, return_complex=False).permute([0, 3, 1, 2])
-        final_output = permuted_stft_output.reshape([*batch_dimensions, channel_dim, 2, -1, permuted_stft_output.shape[-1]]).reshape([*batch_dimensions, channel_dim * 2, -1, permuted_stft_output.shape[-1]])
-
-        if is_non_standard_device: final_output = final_output.to(self.device)
-        return final_output[..., : self.dim_f, :]
-
-    def pad_frequency_dimension(self, input_tensor, batch_dimensions, channel_dim, freq_dim, time_dim, num_freq_bins):
-        return torch.cat([input_tensor, torch.zeros([*batch_dimensions, channel_dim, num_freq_bins - freq_dim, time_dim]).to(input_tensor.device)], -2)
-
-    def calculate_inverse_dimensions(self, input_tensor):
-        channel_dim, freq_dim, time_dim = input_tensor.shape[-3:]
-
-        return input_tensor.shape[:-3], channel_dim, freq_dim, time_dim, self.n_fft // 2 + 1
-
-    def prepare_for_istft(self, padded_tensor, batch_dimensions, channel_dim, num_freq_bins, time_dim):
-        permuted_tensor = padded_tensor.reshape([*batch_dimensions, channel_dim // 2, 2, num_freq_bins, time_dim]).reshape([-1, 2, num_freq_bins, time_dim]).permute([0, 2, 3, 1])
-
-        return permuted_tensor[..., 0] + permuted_tensor[..., 1] * 1.0j
-
-    def inverse(self, input_tensor):
-        is_non_standard_device = not input_tensor.device.type in ["cuda", "cpu"]
-        if is_non_standard_device: input_tensor = input_tensor.cpu()
-
-        batch_dimensions, channel_dim, freq_dim, time_dim, num_freq_bins = self.calculate_inverse_dimensions(input_tensor)
-        final_output = torch.istft(self.prepare_for_istft(self.pad_frequency_dimension(input_tensor, batch_dimensions, channel_dim, freq_dim, time_dim, num_freq_bins), batch_dimensions, channel_dim, num_freq_bins, time_dim), n_fft=self.n_fft, hop_length=self.hop_length, window=self.hann_window.to(input_tensor.device), center=True).reshape([*batch_dimensions, 2, -1])
-
-        if is_non_standard_device: final_output = final_output.to(self.device)
-
-        return final_output

main/library/audioldm2/models.py

@@ -1,330 +0,0 @@
-import os
-import sys
-import torch
-import librosa
-
-import numpy as np
-import torch.nn.functional as F
-
-from scipy.signal import get_window
-from librosa.util import pad_center
-from diffusers import DDIMScheduler, AudioLDM2Pipeline
-from diffusers.models.unets.unet_2d_condition import UNet2DConditionOutput
-from transformers import RobertaTokenizer, RobertaTokenizerFast, VitsTokenizer
-
-sys.path.append(os.getcwd())
-
-from main.configs.config import Config
-from main.library.utils import check_audioldm2
-
-config = Config()
-
-class Pipeline(torch.nn.Module):
-    def __init__(self, model_id, device, double_precision = False, token = None, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.model_id = model_id
-        self.device = device
-        self.double_precision = double_precision
-        self.token = token
-
-    def load_scheduler(self):
-        pass
-
-    def get_melspectrogram(self):
-        pass
-
-    def vae_encode(self, x):
-        pass
-
-    def vae_decode(self, x):
-        pass
-
-    def decode_to_mel(self, x):
-        pass
-
-    def setup_extra_inputs(self, *args, **kwargs):
-        pass
-
-    def encode_text(self, prompts, **kwargs):
-        pass
-
-    def get_variance(self, timestep, prev_timestep):
-        pass
-
-    def get_alpha_prod_t_prev(self, prev_timestep):
-        pass
-
-    def get_noise_shape(self, x0, num_steps):
-        return (num_steps, self.model.unet.config.in_channels, x0.shape[-2], x0.shape[-1])
-
-    def sample_xts_from_x0(self, x0, num_inference_steps = 50):
-        alpha_bar = self.model.scheduler.alphas_cumprod
-        sqrt_one_minus_alpha_bar = (1 - alpha_bar) ** 0.5
-        timesteps = self.model.scheduler.timesteps.to(self.device)
-        t_to_idx = {int(v): k for k, v in enumerate(timesteps)}
-        xts = torch.zeros(self.get_noise_shape(x0, num_inference_steps + 1)).to(x0.device)
-        xts[0] = x0
-
-        for t in reversed(timesteps):
-            idx = num_inference_steps - t_to_idx[int(t)]
-            xts[idx] = x0 * (alpha_bar[t] ** 0.5) + torch.randn_like(x0) * sqrt_one_minus_alpha_bar[t]
-
-        return xts
-
-    def get_zs_from_xts(self, xt, xtm1, noise_pred, t, eta = 0, numerical_fix = True, **kwargs):
-        alpha_bar = self.model.scheduler.alphas_cumprod
-
-        if self.model.scheduler.config.prediction_type == 'epsilon': pred_original_sample = (xt - (1 - alpha_bar[t]) ** 0.5 * noise_pred) / alpha_bar[t] ** 0.5
-        elif self.model.scheduler.config.prediction_type == 'v_prediction': pred_original_sample = (alpha_bar[t] ** 0.5) * xt - ((1 - alpha_bar[t]) ** 0.5) * noise_pred
-
-        prev_timestep = t - self.model.scheduler.config.num_train_timesteps // self.model.scheduler.num_inference_steps
-        alpha_prod_t_prev = self.get_alpha_prod_t_prev(prev_timestep)
-        variance = self.get_variance(t, prev_timestep)
-
-        if self.model.scheduler.config.prediction_type == 'epsilon': radom_noise_pred = noise_pred
-        elif self.model.scheduler.config.prediction_type == 'v_prediction': radom_noise_pred = (alpha_bar[t] ** 0.5) * noise_pred + ((1 - alpha_bar[t]) ** 0.5) * xt
-
-        mu_xt = alpha_prod_t_prev ** (0.5) * pred_original_sample + ((1 - alpha_prod_t_prev - eta * variance) ** (0.5) * radom_noise_pred)
-        z = (xtm1 - mu_xt) / (eta * variance ** 0.5)
-
-        if numerical_fix: xtm1 = mu_xt + (eta * variance ** 0.5)*z
-        return z, xtm1, None
-
-    def reverse_step_with_custom_noise(self, model_output, timestep, sample, variance_noise = None, eta = 0, **kwargs):
-        prev_timestep = timestep - self.model.scheduler.config.num_train_timesteps // self.model.scheduler.num_inference_steps
-        alpha_prod_t = self.model.scheduler.alphas_cumprod[timestep]
-        alpha_prod_t_prev = self.get_alpha_prod_t_prev(prev_timestep)
-        beta_prod_t = 1 - alpha_prod_t
-
-        if self.model.scheduler.config.prediction_type == 'epsilon': pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
-        elif self.model.scheduler.config.prediction_type == 'v_prediction': pred_original_sample = (alpha_prod_t ** 0.5) * sample - (beta_prod_t ** 0.5) * model_output
-
-        variance = self.get_variance(timestep, prev_timestep)
-
-        if self.model.scheduler.config.prediction_type == 'epsilon': model_output_direction = model_output
-        elif self.model.scheduler.config.prediction_type == 'v_prediction': model_output_direction = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample
-
-        prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + ((1 - alpha_prod_t_prev - eta * variance) ** (0.5) * model_output_direction)
-
-        if eta > 0:
-            if variance_noise is None: variance_noise = torch.randn(model_output.shape, device=self.device)
-            prev_sample = prev_sample + (eta * variance ** (0.5) * variance_noise)
-
-        return prev_sample
-
-    def unet_forward(self, sample, timestep, encoder_hidden_states, class_labels = None, timestep_cond = None, attention_mask = None, cross_attention_kwargs = None, added_cond_kwargs = None, down_block_additional_residuals = None, mid_block_additional_residual = None, encoder_attention_mask = None, replace_h_space = None, replace_skip_conns = None, return_dict = True, zero_out_resconns = None):
-        pass
-
-class STFT(torch.nn.Module):
-    def __init__(self, fft_size, hop_size, window_size, window_type="hann"):
-        super().__init__()
-        self.fft_size = fft_size
-        self.hop_size = hop_size
-        self.window_size = window_size
-        self.window_type = window_type
-        
-        scale = fft_size / hop_size
-        fourier_basis = np.fft.fft(np.eye(fft_size))
-
-        cutoff = fft_size // 2 + 1
-        fourier_basis = np.vstack([np.real(fourier_basis[:cutoff, :]), np.imag(fourier_basis[:cutoff, :])])
-        
-        self.forward_basis = torch.FloatTensor(fourier_basis[:, None, :])
-        self.inverse_basis = torch.FloatTensor(np.linalg.pinv(scale * fourier_basis).T[:, None, :])
-        
-        if window_type:
-            assert fft_size >= window_size
-
-            fft_window = torch.from_numpy(pad_center(get_window(window_type, window_size, fftbins=True), size=fft_size)).float()
-            self.forward_basis *= fft_window
-            self.inverse_basis *= fft_window
-        
-        if not hasattr(self, "forward_basis"): self.register_buffer("forward_basis", self.forward_basis)
-        if not hasattr(self, "inverse_basis"): self.register_buffer("inverse_basis", self.inverse_basis)
-    
-    def transform(self, signal):
-        batch_size, num_samples = signal.shape
-        transformed_signal = F.conv1d(F.pad(signal.view(batch_size, 1, num_samples).unsqueeze(1), (self.fft_size // 2, self.fft_size // 2, 0, 0), mode="reflect").squeeze(1), self.forward_basis, stride=self.hop_size, padding=0).cpu()
-        
-        cutoff = self.fft_size // 2 + 1
-        real_part, imag_part = transformed_signal[:, :cutoff, :], transformed_signal[:, cutoff:, :]
-
-        return torch.sqrt(real_part ** 2 + imag_part ** 2), torch.atan2(imag_part, real_part)
-
-class MelSpectrogramProcessor(torch.nn.Module):
-    def __init__(self, fft_size, hop_size, window_size, num_mel_bins, sample_rate, fmin, fmax):
-        super().__init__()
-        self.num_mel_bins = num_mel_bins
-        self.sample_rate = sample_rate
-        self.stft_processor = STFT(fft_size, hop_size, window_size)
-        self.register_buffer("mel_filter", torch.from_numpy(librosa.filters.mel(sr=sample_rate, n_fft=fft_size, n_mels=num_mel_bins, fmin=fmin, fmax=fmax)).float())
-    
-    def compute_mel_spectrogram(self, waveform, normalization_fn=torch.log):
-        assert torch.min(waveform) >= -1
-        assert torch.max(waveform) <= 1
-        
-        magnitudes, _ = self.stft_processor.transform(waveform)
-        return normalization_fn(torch.clamp(torch.matmul(self.mel_filter, magnitudes), min=1e-5))
-
-class AudioLDM2(Pipeline):
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.model = AudioLDM2Pipeline.from_pretrained(self.model_id, local_files_only=True, torch_dtype=torch.float16 if config.is_half else torch.float32).to(self.device)
-
-    def load_scheduler(self):
-        self.model.scheduler = DDIMScheduler.from_pretrained(self.model_id, subfolder="scheduler")
-
-    def get_melspectrogram(self):
-        return MelSpectrogramProcessor(fft_size=1024, hop_size=160, window_size=1024, num_mel_bins=64, sample_rate=16000, fmin=0, fmax=8000)
-
-    def vae_encode(self, x):
-        if x.shape[2] % 4: x = F.pad(x, (0, 0, 4 - (x.shape[2] % 4), 0))
-        output = (self.model.vae.encode(x.half() if config.is_half else x.float()).latent_dist.mode() * self.model.vae.config.scaling_factor)
-        return output.half() if config.is_half else output.float()
-
-    def vae_decode(self, x):
-        return self.model.vae.decode(1 / self.model.vae.config.scaling_factor * x).sample
-
-    def decode_to_mel(self, x):
-        tmp = self.model.mel_spectrogram_to_waveform(x[:, 0].detach().to(torch.float16 if config.is_half else torch.float32)).detach()
-
-        if len(tmp.shape) == 1: tmp = tmp.unsqueeze(0)
-        return tmp
-
-    def encode_text(self, prompts, negative = False, save_compute = False, cond_length = 0, **kwargs):
-        tokenizers, text_encoders = [self.model.tokenizer, self.model.tokenizer_2], [self.model.text_encoder, self.model.text_encoder_2]
-        prompt_embeds_list, attention_mask_list = [], []
-
-        for tokenizer, text_encoder in zip(tokenizers, text_encoders):
-            text_inputs = tokenizer(prompts, padding="max_length" if (save_compute and negative) or isinstance(tokenizer, (RobertaTokenizer, RobertaTokenizerFast)) else True, max_length=tokenizer.model_max_length if (not save_compute) or ((not negative) or isinstance(tokenizer, (RobertaTokenizer, RobertaTokenizerFast, VitsTokenizer))) else cond_length, truncation=True, return_tensors="pt")
-            text_input_ids = text_inputs.input_ids
-
-            attention_mask = text_inputs.attention_mask
-            untruncated_ids = tokenizer(prompts, padding="longest", return_tensors="pt").input_ids
-
-            if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1: -1])
-
-            text_input_ids = text_input_ids.to(self.device)
-            attention_mask = attention_mask.to(self.device)
-
-            with torch.no_grad():
-                if text_encoder.config.model_type == "clap":
-                    prompt_embeds = text_encoder.get_text_features(text_input_ids, attention_mask=attention_mask)
-                    prompt_embeds = prompt_embeds[:, None, :]
-                    attention_mask = attention_mask.new_ones((len(prompts), 1))
-                else: prompt_embeds = text_encoder(text_input_ids, attention_mask=attention_mask)[0]
-
-            prompt_embeds_list.append(prompt_embeds)
-            attention_mask_list.append(attention_mask)
-
-        projection_output = self.model.projection_model(hidden_states=prompt_embeds_list[0], hidden_states_1=prompt_embeds_list[1], attention_mask=attention_mask_list[0], attention_mask_1=attention_mask_list[1])
-        generated_prompt_embeds = self.model.generate_language_model(projection_output.hidden_states, attention_mask=projection_output.attention_mask, max_new_tokens=None)
-        prompt_embeds = prompt_embeds.to(dtype=self.model.text_encoder_2.dtype, device=self.device)
-        return generated_prompt_embeds.to(dtype=self.model.language_model.dtype, device=self.device), prompt_embeds, (attention_mask.to(device=self.device) if attention_mask is not None else torch.ones(prompt_embeds.shape[:2], dtype=torch.long, device=self.device))
-
-    def get_variance(self, timestep, prev_timestep):
-        alpha_prod_t = self.model.scheduler.alphas_cumprod[timestep]
-        alpha_prod_t_prev = self.get_alpha_prod_t_prev(prev_timestep)
-        return ((1 - alpha_prod_t_prev) / (1 - alpha_prod_t)) * (1 - alpha_prod_t / alpha_prod_t_prev)
-
-    def get_alpha_prod_t_prev(self, prev_timestep):
-        return self.model.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.model.scheduler.final_alpha_cumprod
-
-    def unet_forward(self, sample, timestep, encoder_hidden_states, timestep_cond = None, class_labels = None, attention_mask = None, encoder_attention_mask = None, return_dict = True, cross_attention_kwargs = None, mid_block_additional_residual = None, replace_h_space = None, replace_skip_conns = None, zero_out_resconns = None):
-        encoder_hidden_states_1 = class_labels
-        class_labels = None
-        encoder_attention_mask_1 = encoder_attention_mask
-        encoder_attention_mask = None
-        default_overall_up_factor = 2 ** self.model.unet.num_upsamplers
-        forward_upsample_size = False
-        upsample_size = None
-
-        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]): forward_upsample_size = True
-
-        if attention_mask is not None:
-            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
-            attention_mask = attention_mask.unsqueeze(1)
-
-        if encoder_attention_mask is not None:
-            encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
-            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
-
-        if encoder_attention_mask_1 is not None:
-            encoder_attention_mask_1 = (1 - encoder_attention_mask_1.to(sample.dtype)) * -10000.0
-            encoder_attention_mask_1 = encoder_attention_mask_1.unsqueeze(1)
-
-        timesteps = timestep
-        if not torch.is_tensor(timesteps):
-            is_mps = sample.device.type == "mps"
-
-            dtype = (torch.float16 if is_mps else torch.float32) if isinstance(timestep, float) else (torch.int16 if is_mps else torch.int32)
-
-            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
-        elif len(timesteps.shape) == 0: timesteps = timesteps[None].to(sample.device)
-
-        emb = self.model.unet.time_embedding(self.model.unet.time_proj(timesteps.expand(sample.shape[0])).to(dtype=sample.dtype), timestep_cond)
-        aug_emb = None
-
-        if self.model.unet.class_embedding is not None:
-            if class_labels is None: raise ValueError
-
-            if self.model.unet.config.class_embed_type == "timestep": class_labels = self.model.unet.time_proj(class_labels).to(dtype=sample.dtype)
-            class_emb = self.model.unet.class_embedding(class_labels).to(dtype=sample.dtype)
-
-            if self.model.unet.config.class_embeddings_concat: emb = torch.cat([emb, class_emb], dim=-1)
-            else: emb = emb + class_emb
-
-        emb = emb + aug_emb if aug_emb is not None else emb
-        if self.model.unet.time_embed_act is not None: emb = self.model.unet.time_embed_act(emb)
-
-        sample = self.model.unet.conv_in(sample)
-        down_block_res_samples = (sample,)
-
-        for downsample_block in self.model.unet.down_blocks:
-            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention: sample, res_samples = downsample_block(hidden_states=sample, temb=emb, encoder_hidden_states=encoder_hidden_states, attention_mask=attention_mask, cross_attention_kwargs=cross_attention_kwargs, encoder_attention_mask=encoder_attention_mask, encoder_hidden_states_1=encoder_hidden_states_1, encoder_attention_mask_1=encoder_attention_mask_1)
-            else: sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
-
-            down_block_res_samples += res_samples
-
-        if self.model.unet.mid_block is not None: sample = self.model.unet.mid_block(sample, emb, encoder_hidden_states=encoder_hidden_states, attention_mask=attention_mask, cross_attention_kwargs=cross_attention_kwargs, encoder_attention_mask=encoder_attention_mask, encoder_hidden_states_1=encoder_hidden_states_1, encoder_attention_mask_1=encoder_attention_mask_1)
-
-        if replace_h_space is None: h_space = sample.clone()
-        else:
-            h_space = replace_h_space
-            sample = replace_h_space.clone()
-
-        if mid_block_additional_residual is not None: sample = sample + mid_block_additional_residual
-        extracted_res_conns = {}
-
-        for i, upsample_block in enumerate(self.model.unet.up_blocks):
-            is_final_block = i == len(self.model.unet.up_blocks) - 1
-            res_samples = down_block_res_samples[-len(upsample_block.resnets):]
-            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
-
-            if replace_skip_conns is not None and replace_skip_conns.get(i): res_samples = replace_skip_conns.get(i)
-
-            if zero_out_resconns is not None:
-                if (type(zero_out_resconns) is int and i >= (zero_out_resconns - 1)) or type(zero_out_resconns) is list and i in zero_out_resconns: res_samples = [torch.zeros_like(x) for x in res_samples]
-
-            extracted_res_conns[i] = res_samples
-            if not is_final_block and forward_upsample_size: upsample_size = down_block_res_samples[-1].shape[2:]
-
-            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention: sample = upsample_block(hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, encoder_hidden_states=encoder_hidden_states, cross_attention_kwargs=cross_attention_kwargs, upsample_size=upsample_size, attention_mask=attention_mask, encoder_attention_mask=encoder_attention_mask, encoder_hidden_states_1=encoder_hidden_states_1, encoder_attention_mask_1=encoder_attention_mask_1)
-            else: sample = upsample_block(hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size)
-
-        if self.model.unet.conv_norm_out: sample = self.model.unet.conv_act(self.model.unet.conv_norm_out(sample))
-        sample = self.model.unet.conv_out(sample)
-
-        if not return_dict: return (sample,)
-        return UNet2DConditionOutput(sample=sample), h_space, extracted_res_conns
-
-def load_model(model, device):
-    check_audioldm2(model)
-
-    ldm_stable = AudioLDM2(model_id=os.path.join("assets", "models", "audioldm2", model), device=device, double_precision=False)
-    ldm_stable.load_scheduler()
-
-    if torch.cuda.is_available(): torch.cuda.empty_cache()
-    elif torch.backends.mps.is_available(): torch.mps.empty_cache()
-
-    return ldm_stable

main/library/audioldm2/utils.py

@@ -1,40 +0,0 @@
-import torch
-import librosa
-import torchaudio
-
-import numpy as np
-
-def compute_mel_spectrogram(audio, stft_processor):
-    return stft_processor.compute_mel_spectrogram(torch.autograd.Variable(torch.clip(torch.FloatTensor(audio).unsqueeze(0), -1, 1), requires_grad=False)).squeeze(0).numpy().astype(np.float32)
-
-def pad_spectrogram(spectrogram, target_length=1024):
-    pad_amount = target_length - spectrogram.shape[0]
-    spectrogram = torch.nn.functional.pad(spectrogram, (0, 0, 0, pad_amount)) if pad_amount > 0 else spectrogram[:target_length, :]
-    
-    if spectrogram.size(-1) % 2 != 0: spectrogram = spectrogram[..., :-1]
-    return spectrogram
-
-def pad_waveform(waveform, segment_length):
-    waveform_length = waveform.shape[-1]
-    assert waveform_length > 100
-    
-    if segment_length is None or waveform_length == segment_length: return waveform
-    elif waveform_length > segment_length: return waveform[:, :segment_length]
-    
-    padded_waveform = np.zeros((1, segment_length))
-    padded_waveform[:, :waveform_length] = waveform
-    return padded_waveform
-
-def normalize(waveform):
-    waveform -= np.mean(waveform)
-    return (waveform / (np.max(np.abs(waveform)) + 1e-8)) * 0.5
-
-def process_audio(y, sr, segment_length):
-    normalized_waveform = normalize(torchaudio.functional.resample(torch.from_numpy(y), orig_freq=sr, new_freq=16000).numpy())[None, ...]
-    return 0.5 * (pad_waveform(normalized_waveform, segment_length) / np.max(np.abs(normalized_waveform)))
-
-def load_audio(audio_path, stft_processor, device=None):
-    y, sr = librosa.load(audio_path, sr=None)
-    duration = len(y) / sr
-    
-    return pad_spectrogram(torch.FloatTensor(compute_mel_spectrogram(torch.FloatTensor(process_audio(y, sr, int(duration * 102.4) * 160)[0, ...]), stft_processor).T), int(duration * 102.4)).unsqueeze(0).unsqueeze(0).to(device), duration

main/library/predictors/CREPE.py

@@ -1,210 +0,0 @@
-import os
-import torch
-import librosa
-import functools
-import scipy.stats
-
-import numpy as np
-
-CENTS_PER_BIN, MAX_FMAX, PITCH_BINS, SAMPLE_RATE, WINDOW_SIZE = 20, 2006, 360, 16000, 1024  
-
-class Crepe(torch.nn.Module):
-    def __init__(self, model='full'):
-        super().__init__()
-        if model == 'full':
-            in_channels = [1, 1024, 128, 128, 128, 256]
-            out_channels = [1024, 128, 128, 128, 256, 512]
-            self.in_features = 2048
-        elif model == 'large':
-            in_channels = [1, 768, 96, 96, 96, 192]
-            out_channels = [768, 96, 96, 96, 192, 384]
-            self.in_features = 1536
-        elif model == 'medium':
-            in_channels = [1, 512, 64, 64, 64, 128]
-            out_channels = [512, 64, 64, 64, 128, 256]
-            self.in_features = 1024
-        elif model == 'small':
-            in_channels = [1, 256, 32, 32, 32, 64]
-            out_channels = [256, 32, 32, 32, 64, 128]
-            self.in_features = 512
-        elif model == 'tiny':
-            in_channels = [1, 128, 16, 16, 16, 32]
-            out_channels = [128, 16, 16, 16, 32, 64]
-            self.in_features = 256
-
-        kernel_sizes = [(512, 1)] + 5 * [(64, 1)]
-        strides = [(4, 1)] + 5 * [(1, 1)]
-
-        batch_norm_fn = functools.partial(torch.nn.BatchNorm2d, eps=0.0010000000474974513, momentum=0.0)
-
-        self.conv1 = torch.nn.Conv2d(in_channels=in_channels[0], out_channels=out_channels[0], kernel_size=kernel_sizes[0], stride=strides[0])
-        self.conv1_BN = batch_norm_fn(num_features=out_channels[0])
-        self.conv2 = torch.nn.Conv2d(in_channels=in_channels[1], out_channels=out_channels[1], kernel_size=kernel_sizes[1], stride=strides[1])
-        self.conv2_BN = batch_norm_fn(num_features=out_channels[1])
-
-        self.conv3 = torch.nn.Conv2d(in_channels=in_channels[2], out_channels=out_channels[2], kernel_size=kernel_sizes[2], stride=strides[2])
-        self.conv3_BN = batch_norm_fn(num_features=out_channels[2])
-        self.conv4 = torch.nn.Conv2d(in_channels=in_channels[3], out_channels=out_channels[3], kernel_size=kernel_sizes[3], stride=strides[3])
-        self.conv4_BN = batch_norm_fn(num_features=out_channels[3])
-
-        self.conv5 = torch.nn.Conv2d(in_channels=in_channels[4], out_channels=out_channels[4], kernel_size=kernel_sizes[4], stride=strides[4])
-        self.conv5_BN = batch_norm_fn(num_features=out_channels[4])
-        self.conv6 = torch.nn.Conv2d(in_channels=in_channels[5], out_channels=out_channels[5], kernel_size=kernel_sizes[5], stride=strides[5])
-        self.conv6_BN = batch_norm_fn(num_features=out_channels[5])
-
-        self.classifier = torch.nn.Linear(in_features=self.in_features, out_features=PITCH_BINS)
-
-    def forward(self, x, embed=False):
-        x = self.embed(x)
-        if embed: return x
-
-        return torch.sigmoid(self.classifier(self.layer(x, self.conv6, self.conv6_BN).permute(0, 2, 1, 3).reshape(-1, self.in_features)))
-
-    def embed(self, x):
-        x = x[:, None, :, None]
-
-        return self.layer(self.layer(self.layer(self.layer(self.layer(x, self.conv1, self.conv1_BN, (0, 0, 254, 254)), self.conv2, self.conv2_BN), self.conv3, self.conv3_BN), self.conv4, self.conv4_BN), self.conv5, self.conv5_BN)
-
-    def layer(self, x, conv, batch_norm, padding=(0, 0, 31, 32)):
-        return torch.nn.functional.max_pool2d(batch_norm(torch.nn.functional.relu(conv(torch.nn.functional.pad(x, padding)))), (2, 1), (2, 1))
-
-def viterbi(logits):
-    if not hasattr(viterbi, 'transition'):
-        xx, yy = np.meshgrid(range(360), range(360))
-        transition = np.maximum(12 - abs(xx - yy), 0)
-        viterbi.transition = transition / transition.sum(axis=1, keepdims=True)
-
-    with torch.no_grad():
-        probs = torch.nn.functional.softmax(logits, dim=1)
-
-    bins = torch.tensor(np.array([librosa.sequence.viterbi(sequence, viterbi.transition).astype(np.int64) for sequence in probs.cpu().numpy()]), device=probs.device)
-    return bins, bins_to_frequency(bins)
-
-def predict(audio, sample_rate, hop_length=None, fmin=50, fmax=MAX_FMAX, model='full', return_periodicity=False, batch_size=None, device='cpu', pad=True, providers=None, onnx=False):
-    results = []
-
-    if onnx:
-        import onnxruntime as ort
-
-        sess_options = ort.SessionOptions()
-        sess_options.log_severity_level = 3
-
-        session = ort.InferenceSession(os.path.join("assets", "models", "predictors", f"crepe_{model}.onnx"), sess_options=sess_options, providers=providers)
-
-        for frames in preprocess(audio, sample_rate, hop_length, batch_size, device, pad):
-            result = postprocess(torch.tensor(session.run([session.get_outputs()[0].name], {session.get_inputs()[0].name: frames.cpu().numpy()})[0].transpose(1, 0)[None]), fmin, fmax, return_periodicity)
-            results.append((result[0], result[1]) if isinstance(result, tuple) else result)
-
-        del session
-
-        if return_periodicity:
-            pitch, periodicity = zip(*results)
-            return torch.cat(pitch, 1), torch.cat(periodicity, 1)
-
-        return torch.cat(results, 1)
-    else:
-        with torch.no_grad():
-            for frames in preprocess(audio, sample_rate, hop_length, batch_size, device, pad):
-                result = postprocess(infer(frames, model, device, embed=False).reshape(audio.size(0), -1, PITCH_BINS).transpose(1, 2), fmin, fmax, return_periodicity)
-                results.append((result[0].to(audio.device), result[1].to(audio.device)) if isinstance(result, tuple) else result.to(audio.device))
-
-        if return_periodicity:
-            pitch, periodicity = zip(*results)
-            return torch.cat(pitch, 1), torch.cat(periodicity, 1)
-        
-        return torch.cat(results, 1)
-
-def bins_to_frequency(bins):
-    cents = CENTS_PER_BIN * bins + 1997.3794084376191
-    return 10 * 2 ** ((cents + cents.new_tensor(scipy.stats.triang.rvs(c=0.5, loc=-CENTS_PER_BIN, scale=2 * CENTS_PER_BIN, size=cents.size()))) / 1200)
-
-def frequency_to_bins(frequency, quantize_fn=torch.floor):
-    return quantize_fn(((1200 * torch.log2(frequency / 10)) - 1997.3794084376191) / CENTS_PER_BIN).int()
-
-def infer(frames, model='full', device='cpu', embed=False):
-    if not hasattr(infer, 'model') or not hasattr(infer, 'capacity') or (hasattr(infer, 'capacity') and infer.capacity != model): load_model(device, model)
-    infer.model = infer.model.to(device)
-
-    return infer.model(frames, embed=embed)
-
-def load_model(device, capacity='full'):
-    infer.capacity = capacity
-    infer.model = Crepe(capacity)
-    infer.model.load_state_dict(torch.load(os.path.join("assets", "models", "predictors", f"crepe_{capacity}.pth"), map_location=device))
-    infer.model = infer.model.to(torch.device(device))
-    infer.model.eval()
-
-def postprocess(probabilities, fmin=0, fmax=MAX_FMAX, return_periodicity=False):
-    probabilities = probabilities.detach()
-
-    probabilities[:, :frequency_to_bins(torch.tensor(fmin))] = -float('inf')
-    probabilities[:, frequency_to_bins(torch.tensor(fmax), torch.ceil):] = -float('inf')
-
-    bins, pitch = viterbi(probabilities)
-
-    if not return_periodicity: return pitch
-    return pitch, periodicity(probabilities, bins)
-
-def preprocess(audio, sample_rate, hop_length=None, batch_size=None, device='cpu', pad=True):
-    hop_length = sample_rate // 100 if hop_length is None else hop_length
-
-    if sample_rate != SAMPLE_RATE:
-        audio = torch.tensor(librosa.resample(audio.detach().cpu().numpy().squeeze(0), orig_sr=sample_rate, target_sr=SAMPLE_RATE, res_type="soxr_vhq"), device=audio.device).unsqueeze(0)
-        hop_length = int(hop_length * SAMPLE_RATE / sample_rate)
-
-    if pad:
-        total_frames = 1 + int(audio.size(1) // hop_length)
-        audio = torch.nn.functional.pad(audio, (WINDOW_SIZE // 2, WINDOW_SIZE // 2))
-    else: total_frames = 1 + int((audio.size(1) - WINDOW_SIZE) // hop_length)
-
-    batch_size = total_frames if batch_size is None else batch_size
-
-    for i in range(0, total_frames, batch_size):
-        frames = torch.nn.functional.unfold(audio[:, None, None, max(0, i * hop_length):min(audio.size(1), (i + batch_size - 1) * hop_length + WINDOW_SIZE)], kernel_size=(1, WINDOW_SIZE), stride=(1, hop_length))
-        frames = frames.transpose(1, 2).reshape(-1, WINDOW_SIZE).to(device)
-        frames -= frames.mean(dim=1, keepdim=True)
-        frames /= torch.max(torch.tensor(1e-10, device=frames.device), frames.std(dim=1, keepdim=True))
-
-        yield frames
-
-def periodicity(probabilities, bins):
-    probs_stacked = probabilities.transpose(1, 2).reshape(-1, PITCH_BINS)
-    periodicity = probs_stacked.gather(1, bins.reshape(-1, 1).to(torch.int64))
-    
-    return periodicity.reshape(probabilities.size(0), probabilities.size(2))
-
-def mean(signals, win_length=9):
-    assert signals.dim() == 2
-
-    signals = signals.unsqueeze(1)
-    mask = ~torch.isnan(signals)
-    padding = win_length // 2
-
-    ones_kernel = torch.ones(signals.size(1), 1, win_length, device=signals.device)
-    avg_pooled = torch.nn.functional.conv1d(torch.where(mask, signals, torch.zeros_like(signals)), ones_kernel, stride=1, padding=padding) / torch.nn.functional.conv1d(mask.float(), ones_kernel, stride=1, padding=padding).clamp(min=1) 
-    avg_pooled[avg_pooled == 0] = float("nan")
-
-    return avg_pooled.squeeze(1)
-
-def median(signals, win_length):
-    assert signals.dim() == 2
-
-    signals = signals.unsqueeze(1)
-    mask = ~torch.isnan(signals)
-    padding = win_length // 2
-
-    x = torch.nn.functional.pad(torch.where(mask, signals, torch.zeros_like(signals)), (padding, padding), mode="reflect")
-    mask = torch.nn.functional.pad(mask.float(), (padding, padding), mode="constant", value=0)
-
-    x = x.unfold(2, win_length, 1)
-    mask = mask.unfold(2, win_length, 1)
-
-    x = x.contiguous().view(x.size()[:3] + (-1,))
-    mask = mask.contiguous().view(mask.size()[:3] + (-1,))
-
-    x_sorted, _ = torch.sort(torch.where(mask.bool(), x.float(), float("inf")).to(x), dim=-1)
-
-    median_pooled = x_sorted.gather(-1, ((mask.sum(dim=-1) - 1) // 2).clamp(min=0).unsqueeze(-1).long()).squeeze(-1)
-    median_pooled[torch.isinf(median_pooled)] = float("nan")
-
-    return median_pooled.squeeze(1)

main/library/predictors/FCPE.py

@@ -1,1000 +0,0 @@
-import os
-import math
-import torch
-
-import numpy as np
-import onnxruntime as ort
-import torch.nn.functional as F
-
-from torch import nn, einsum
-from functools import partial
-from torchaudio.transforms import Resample
-from einops import rearrange, repeat, pack, unpack
-from torch.nn.utils.parametrizations import weight_norm
-
-from librosa.filters import mel as librosa_mel_fn
-
-os.environ["LRU_CACHE_CAPACITY"] = "3"
-
-def exists(val):
-    return val is not None
-
-def default(value, d):
-    return value if exists(value) else d
-
-def empty(tensor):
-    return tensor.numel() == 0
-
-def decrypt_model(input_path):
-    from io import BytesIO
-    from Crypto.Cipher import AES
-    from Crypto.Util.Padding import unpad
-
-    with open(input_path, "rb") as f:
-        data = f.read()
-
-    with open(os.path.join("main", "configs", "decrypt.bin"), "rb") as f:
-        key = f.read()
-
-    return BytesIO(unpad(AES.new(key, AES.MODE_CBC, data[:16]).decrypt(data[16:]), AES.block_size)).read()
-
-def l2_regularization(model, l2_alpha):
-    l2_loss = []
-    for module in model.modules():
-        if type(module) is nn.Conv2d: l2_loss.append((module.weight**2).sum() / 2.0)
-
-    return l2_alpha * sum(l2_loss)
-
-def pad_to_multiple(tensor, multiple, dim=-1, value=0):
-    seqlen = tensor.shape[dim]
-    m = seqlen / multiple
-    if m.is_integer(): return False, tensor
-    return True, F.pad(tensor, (*((0,) * (-1 - dim) * 2), 0, (math.ceil(m) * multiple - seqlen)), value = value)
-
-def look_around(x, backward = 1, forward = 0, pad_value = -1, dim = 2):
-    t = x.shape[1]
-    dims = (len(x.shape) - dim) * (0, 0)
-    padded_x = F.pad(x, (*dims, backward, forward), value = pad_value)
-    return torch.cat([padded_x[:, ind:(ind + t), ...] for ind in range(forward + backward + 1)], dim = dim)
-
-def rotate_half(x):
-    x1, x2 = rearrange(x, 'b ... (r d) -> b ... r d', r = 2).unbind(dim = -2)
-    return torch.cat((-x2, x1), dim = -1)
-
-def apply_rotary_pos_emb(q, k, freqs, scale = 1):
-    q_len = q.shape[-2]
-    q_freqs = freqs[..., -q_len:, :]
-    inv_scale = scale ** -1
-    if scale.ndim == 2: scale = scale[-q_len:, :]
-    q = (q * q_freqs.cos() * scale) + (rotate_half(q) * q_freqs.sin() * scale)
-    k = (k * freqs.cos() * inv_scale) + (rotate_half(k) * freqs.sin() * inv_scale)
-
-    return q, k
-
-def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
-    return torch.log(torch.clamp(x, min=clip_val) * C)
-
-def orthogonal_matrix_chunk(cols, qr_uniform_q=False, device=None):
-    unstructured_block = torch.randn((cols, cols), device=device)
-    q, r = torch.linalg.qr(unstructured_block.cpu(), mode="reduced")
-    q, r = map(lambda t: t.to(device), (q, r))
-    if qr_uniform_q:
-        d = torch.diag(r, 0)
-        q *= d.sign()
-
-    return q.t()
-
-def linear_attention(q, k, v):
-    return einsum("...ed,...nd->...ne", k, q) if v is None else einsum("...de,...nd,...n->...ne", einsum("...nd,...ne->...de", k, v), q, 1.0 / (einsum("...nd,...d->...n", q, k.sum(dim=-2).type_as(q)) + 1e-8))
-
-def gaussian_orthogonal_random_matrix(nb_rows, nb_columns, scaling=0, qr_uniform_q=False, device=None):
-    nb_full_blocks = int(nb_rows / nb_columns)
-    block_list = []
-    for _ in range(nb_full_blocks):
-        block_list.append(orthogonal_matrix_chunk(nb_columns, qr_uniform_q=qr_uniform_q, device=device))
-
-    remaining_rows = nb_rows - nb_full_blocks * nb_columns
-    if remaining_rows > 0: block_list.append(orthogonal_matrix_chunk(nb_columns, qr_uniform_q=qr_uniform_q, device=device)[:remaining_rows])
-    if scaling == 0: multiplier = torch.randn((nb_rows, nb_columns), device=device).norm(dim=1)
-    elif scaling == 1: multiplier = math.sqrt((float(nb_columns))) * torch.ones((nb_rows,), device=device)
-    else: raise ValueError(f"{scaling} != 0, 1")
-
-    return torch.diag(multiplier) @ torch.cat(block_list)
-
-def calc_same_padding(kernel_size):
-    pad = kernel_size // 2
-    return (pad, pad - (kernel_size + 1) % 2)
-
-def softmax_kernel(data, *, projection_matrix, is_query, normalize_data=True, eps=1e-4, device=None):
-    b, h, *_ = data.shape
-    
-    data_normalizer = (data.shape[-1] ** -0.25) if normalize_data else 1.0
-    ratio = projection_matrix.shape[0] ** -0.5
-    data_dash = torch.einsum("...id,...jd->...ij", (data_normalizer * data), repeat(projection_matrix, "j d -> b h j d", b=b, h=h).type_as(data))
-    diag_data = ((torch.sum(data**2, dim=-1) / 2.0) * (data_normalizer**2)).unsqueeze(dim=-1)
-
-    return (ratio * (torch.exp(data_dash - diag_data - torch.max(data_dash, dim=-1, keepdim=True).values) + eps) if is_query else ratio * (torch.exp(data_dash - diag_data + eps))).type_as(data)
-
-def torch_interp(x, xp, fp):
-    sort_idx = torch.argsort(xp)
-    xp = xp[sort_idx]
-    fp = fp[sort_idx]
-
-    right_idxs = torch.searchsorted(xp, x).clamp(max=len(xp) - 1)
-    left_idxs = (right_idxs - 1).clamp(min=0)
-    x_left = xp[left_idxs]
-    y_left = fp[left_idxs]
-
-    interp_vals = y_left + ((x - x_left) * (fp[right_idxs] - y_left) / (xp[right_idxs] - x_left))
-    interp_vals[x < xp[0]] = fp[0]
-    interp_vals[x > xp[-1]] = fp[-1]
-
-    return interp_vals
-
-def batch_interp_with_replacement_detach(uv, f0):
-    result = f0.clone()
-    for i in range(uv.shape[0]):
-        interp_vals = torch_interp(torch.where(uv[i])[-1], torch.where(~uv[i])[-1], f0[i][~uv[i]]).detach()
-        result[i][uv[i]] = interp_vals
-        
-    return result
-
-def catch_none_args_must(x, func_name, warning_str):
-    if x is None: raise ValueError(f'[Error] {warning_str}\n[Error] > {func_name}')
-    else: return x
-
-def catch_none_args_opti(x, default, func_name, warning_str=None, level='WARN'):
-    return default if x is None else x
-
-def spawn_wav2mel(args, device = None):
-    _type = args.mel.type
-    if (str(_type).lower() == 'none') or (str(_type).lower() == 'default'): _type = 'default'
-    elif str(_type).lower() == 'stft': _type = 'stft'
-    wav2mel = Wav2MelModule(sr=catch_none_args_opti(args.mel.sr, default=16000, func_name='torchfcpe.tools.spawn_wav2mel', warning_str='args.mel.sr is None'), n_mels=catch_none_args_opti(args.mel.num_mels, default=128, func_name='torchfcpe.tools.spawn_wav2mel', warning_str='args.mel.num_mels is None'), n_fft=catch_none_args_opti(args.mel.n_fft, default=1024, func_name='torchfcpe.tools.spawn_wav2mel', warning_str='args.mel.n_fft is None'), win_size=catch_none_args_opti(args.mel.win_size, default=1024, func_name='torchfcpe.tools.spawn_wav2mel', warning_str='args.mel.win_size is None'), hop_length=catch_none_args_opti(args.mel.hop_size, default=160, func_name='torchfcpe.tools.spawn_wav2mel', warning_str='args.mel.hop_size is None'), fmin=catch_none_args_opti(args.mel.fmin, default=0, func_name='torchfcpe.tools.spawn_wav2mel', warning_str='args.mel.fmin is None'), fmax=catch_none_args_opti(args.mel.fmax, default=8000, func_name='torchfcpe.tools.spawn_wav2mel', warning_str='args.mel.fmax is None'), clip_val=1e-05, mel_type=_type)
-    device = catch_none_args_opti(device, default='cpu', func_name='torchfcpe.tools.spawn_wav2mel', warning_str='.device is None')
-    
-    return wav2mel.to(torch.device(device))
-
-def ensemble_f0(f0s, key_shift_list, tta_uv_penalty):
-    device = f0s.device
-    f0s = f0s / (torch.pow(2, torch.tensor(key_shift_list, device=device).to(device).unsqueeze(0).unsqueeze(0) / 12))
-    notes = torch.log2(f0s / 440) * 12 + 69
-    notes[notes < 0] = 0
-
-    uv_penalty = tta_uv_penalty**2
-    dp = torch.zeros_like(notes, device=device)
-    backtrack = torch.zeros_like(notes, device=device).long()
-    dp[:, 0, :] = (notes[:, 0, :] <= 0) * uv_penalty
-
-    for t in range(1, notes.size(1)):
-        penalty = torch.zeros([notes.size(0), notes.size(2), notes.size(2)], device=device)
-        t_uv = notes[:, t, :] <= 0
-        penalty += uv_penalty * t_uv.unsqueeze(1)
-
-        t1_uv = notes[:, t - 1, :] <= 0
-        l2 = torch.pow((notes[:, t - 1, :].unsqueeze(-1) - notes[:, t, :].unsqueeze(1)) * (~t1_uv).unsqueeze(-1) * (~t_uv).unsqueeze(1), 2) - 0.5
-        l2 = l2 * (l2 > 0)
-
-        penalty += l2
-        penalty += t1_uv.unsqueeze(-1) * (~t_uv).unsqueeze(1) * uv_penalty * 2
-
-        min_value, min_indices = torch.min(dp[:, t - 1, :].unsqueeze(-1) + penalty, dim=1)
-        dp[:, t, :] = min_value
-        backtrack[:, t, :] = min_indices
-
-    t = f0s.size(1) - 1
-    f0_result = torch.zeros_like(f0s[:, :, 0], device=device)
-    min_indices = torch.argmin(dp[:, t, :], dim=-1)
-
-    for i in range(0, t + 1):
-        f0_result[:, t - i] = f0s[:, t - i, min_indices]
-        min_indices = backtrack[:, t - i, min_indices]
-
-    return f0_result.unsqueeze(-1)
-
-class LocalAttention(nn.Module):
-    def __init__(self, window_size, causal = False, look_backward = 1, look_forward = None, dropout = 0., shared_qk = False, rel_pos_emb_config = None, dim = None, autopad = False, exact_windowsize = False, scale = None, use_rotary_pos_emb = True, use_xpos = False, xpos_scale_base = None):
-        super().__init__()
-        look_forward = default(look_forward, 0 if causal else 1)
-        assert not (causal and look_forward > 0)
-        self.scale = scale
-        self.window_size = window_size
-        self.autopad = autopad
-        self.exact_windowsize = exact_windowsize
-        self.causal = causal
-        self.look_backward = look_backward
-        self.look_forward = look_forward
-        self.dropout = nn.Dropout(dropout)
-        self.shared_qk = shared_qk
-        self.rel_pos = None
-        self.use_xpos = use_xpos
-        if use_rotary_pos_emb and (exists(rel_pos_emb_config) or exists(dim)): 
-            if exists(rel_pos_emb_config): dim = rel_pos_emb_config[0]
-            self.rel_pos = SinusoidalEmbeddings(dim, use_xpos = use_xpos, scale_base = default(xpos_scale_base, window_size // 2))
-
-    def forward(self, q, k, v, mask = None, input_mask = None, attn_bias = None, window_size = None):
-        mask = default(mask, input_mask)
-        assert not (exists(window_size) and not self.use_xpos)
-
-        _, autopad, pad_value, window_size, causal, look_backward, look_forward, shared_qk = q.shape, self.autopad, -1, default(window_size, self.window_size), self.causal, self.look_backward, self.look_forward, self.shared_qk
-        (q, packed_shape), (k, _), (v, _) = map(lambda t: pack([t], '* n d'), (q, k, v))
-
-        if autopad:
-            orig_seq_len = q.shape[1]
-            (_, q), (_, k), (_, v) = map(lambda t: pad_to_multiple(t, self.window_size, dim = -2), (q, k, v))
-
-        b, n, dim_head, device, dtype = *q.shape, q.device, q.dtype
-        scale = default(self.scale, dim_head ** -0.5)
-
-        assert (n % window_size) == 0
-        windows = n // window_size
-
-        if shared_qk: k = F.normalize(k, dim = -1).type(k.dtype)
-
-        seq = torch.arange(n, device = device)
-        b_t = rearrange(seq, '(w n) -> 1 w n', w = windows, n = window_size)
-        bq, bk, bv = map(lambda t: rearrange(t, 'b (w n) d -> b w n d', w = windows), (q, k, v))
-
-        bq = bq * scale
-        look_around_kwargs = dict(backward =  look_backward, forward =  look_forward, pad_value = pad_value)
-
-        bk = look_around(bk, **look_around_kwargs)
-        bv = look_around(bv, **look_around_kwargs)
-
-        if exists(self.rel_pos):
-            pos_emb, xpos_scale = self.rel_pos(bk)
-            bq, bk = apply_rotary_pos_emb(bq, bk, pos_emb, scale = xpos_scale)
-
-        bq_t = b_t
-        bq_k = look_around(b_t, **look_around_kwargs)
-        bq_t = rearrange(bq_t, '... i -> ... i 1')
-        bq_k = rearrange(bq_k, '... j -> ... 1 j')
-
-        pad_mask = bq_k == pad_value
-        sim = einsum('b h i e, b h j e -> b h i j', bq, bk)
-
-        if exists(attn_bias):
-            heads = attn_bias.shape[0]
-            assert (b % heads) == 0
-
-            attn_bias = repeat(attn_bias, 'h i j -> (b h) 1 i j', b = b // heads)
-            sim = sim + attn_bias
-
-        mask_value = -torch.finfo(sim.dtype).max
-        if shared_qk:
-            self_mask = bq_t == bq_k
-            sim = sim.masked_fill(self_mask, -5e4)
-            del self_mask
-
-        if causal:
-            causal_mask = bq_t < bq_k
-            if self.exact_windowsize: causal_mask = causal_mask | (bq_t > (bq_k + (self.window_size * self.look_backward)))
-            sim = sim.masked_fill(causal_mask, mask_value)
-            del causal_mask
-
-        sim = sim.masked_fill(((bq_k - (self.window_size * self.look_forward)) > bq_t) | (bq_t > (bq_k + (self.window_size * self.look_backward))) | pad_mask, mask_value) if not causal and self.exact_windowsize else sim.masked_fill(pad_mask, mask_value)
-
-        if exists(mask):
-            batch = mask.shape[0]
-            assert (b % batch) == 0
-
-            h = b // mask.shape[0]
-            if autopad: _, mask = pad_to_multiple(mask, window_size, dim = -1, value = False)
-
-            mask = repeat(rearrange(look_around(rearrange(mask, '... (w n) -> (...) w n', w = windows, n = window_size), **{**look_around_kwargs, 'pad_value': False}), '... j -> ... 1 j'), 'b ... -> (b h) ...', h = h)
-            sim = sim.masked_fill(~mask, mask_value)
-
-            del mask
-
-        out = rearrange(einsum('b h i j, b h j e -> b h i e', self.dropout(sim.softmax(dim = -1)), bv), 'b w n d -> b (w n) d')
-        if autopad: out = out[:, :orig_seq_len, :]
-
-        out, *_ = unpack(out, packed_shape, '* n d')
-        return out
-    
-class SinusoidalEmbeddings(nn.Module):
-    def __init__(self, dim, scale_base = None, use_xpos = False, theta = 10000):
-        super().__init__()
-        inv_freq = 1. / (theta ** (torch.arange(0, dim, 2).float() / dim))
-        self.register_buffer('inv_freq', inv_freq)
-        self.use_xpos = use_xpos
-        self.scale_base = scale_base
-        assert not (use_xpos and not exists(scale_base))
-        scale = (torch.arange(0, dim, 2) + 0.4 * dim) / (1.4 * dim)
-        self.register_buffer('scale', scale, persistent = False)
-
-    def forward(self, x):
-        seq_len, device = x.shape[-2], x.device
-        t = torch.arange(seq_len, device = x.device).type_as(self.inv_freq)
-
-        freqs = torch.einsum('i , j -> i j', t, self.inv_freq)
-        freqs =  torch.cat((freqs, freqs), dim = -1)
-
-        if not self.use_xpos: return freqs, torch.ones(1, device = device)
-
-        power = (t - (seq_len // 2)) / self.scale_base
-        scale = self.scale ** rearrange(power, 'n -> n 1')
-
-        return freqs, torch.cat((scale, scale), dim = -1)
-
-class STFT:
-    def __init__(self, sr=22050, n_mels=80, n_fft=1024, win_size=1024, hop_length=256, fmin=20, fmax=11025, clip_val=1e-5):
-        self.target_sr = sr
-        self.n_mels = n_mels
-        self.n_fft = n_fft
-        self.win_size = win_size
-        self.hop_length = hop_length
-        self.fmin = fmin
-        self.fmax = fmax
-        self.clip_val = clip_val
-        self.mel_basis = {}
-        self.hann_window = {}
-
-    def get_mel(self, y, keyshift=0, speed=1, center=False, train=False):
-        n_fft = self.n_fft
-        win_size = self.win_size
-        hop_length = self.hop_length
-        fmax = self.fmax
-        factor = 2 ** (keyshift / 12)
-        win_size_new = int(np.round(win_size * factor))
-        hop_length_new = int(np.round(hop_length * speed))
-        mel_basis = self.mel_basis if not train else {}
-        hann_window = self.hann_window if not train else {}
-        mel_basis_key = str(fmax) + "_" + str(y.device)
-
-        if mel_basis_key not in mel_basis: mel_basis[mel_basis_key] = torch.from_numpy(librosa_mel_fn(sr=self.target_sr, n_fft=n_fft, n_mels=self.n_mels, fmin=self.fmin, fmax=fmax)).float().to(y.device)
-        keyshift_key = str(keyshift) + "_" + str(y.device)
-        if keyshift_key not in hann_window: hann_window[keyshift_key] = torch.hann_window(win_size_new).to(y.device)
-
-        pad_left = (win_size_new - hop_length_new) // 2
-        pad_right = max((win_size_new - hop_length_new + 1) // 2, win_size_new - y.size(-1) - pad_left)
-        spec = torch.stft(F.pad(y.unsqueeze(1), (pad_left, pad_right), mode="reflect" if pad_right < y.size(-1) else "constant").squeeze(1), int(np.round(n_fft * factor)), hop_length=hop_length_new, win_length=win_size_new, window=hann_window[keyshift_key], center=center, pad_mode="reflect", normalized=False, onesided=True, return_complex=True)
-        spec = torch.sqrt(spec.real.pow(2) + spec.imag.pow(2) + (1e-9))
-
-        if keyshift != 0:
-            size = n_fft // 2 + 1
-            resize = spec.size(1)
-            spec = (F.pad(spec, (0, 0, 0, size - resize)) if resize < size else spec[:, :size, :]) * win_size / win_size_new
-
-        return dynamic_range_compression_torch(torch.matmul(mel_basis[mel_basis_key], spec), clip_val=self.clip_val)
-
-class PCmer(nn.Module):
-    def __init__(self, num_layers, num_heads, dim_model, dim_keys, dim_values, residual_dropout, attention_dropout):
-        super().__init__()
-        self.num_layers = num_layers
-        self.num_heads = num_heads
-        self.dim_model = dim_model
-        self.dim_values = dim_values
-        self.dim_keys = dim_keys
-        self.residual_dropout = residual_dropout
-        self.attention_dropout = attention_dropout
-        self._layers = nn.ModuleList([_EncoderLayer(self) for _ in range(num_layers)])
-
-    def forward(self, phone, mask=None):
-        for layer in self._layers:
-            phone = layer(phone, mask)
-
-        return phone
-
-class _EncoderLayer(nn.Module):
-    def __init__(self, parent):
-        super().__init__()
-        self.conformer = ConformerConvModule_LEGACY(parent.dim_model)
-        self.norm = nn.LayerNorm(parent.dim_model)
-        self.dropout = nn.Dropout(parent.residual_dropout)
-        self.attn = SelfAttention(dim=parent.dim_model, heads=parent.num_heads, causal=False)
-
-    def forward(self, phone, mask=None):
-        phone = phone + (self.attn(self.norm(phone), mask=mask))
-        return phone + (self.conformer(phone))
-
-class ConformerNaiveEncoder(nn.Module):
-    def __init__(self, num_layers, num_heads, dim_model, use_norm = False, conv_only = False, conv_dropout = 0, atten_dropout = 0):
-        super().__init__()
-        self.num_layers = num_layers
-        self.num_heads = num_heads
-        self.dim_model = dim_model
-        self.use_norm = use_norm
-        self.residual_dropout = 0.1  
-        self.attention_dropout = 0.1  
-        self.encoder_layers = nn.ModuleList([CFNEncoderLayer(dim_model, num_heads, use_norm, conv_only, conv_dropout, atten_dropout) for _ in range(num_layers)])
-
-    def forward(self, x, mask=None):
-        for (_, layer) in enumerate(self.encoder_layers):
-            x = layer(x, mask)
-
-        return x 
-
-class CFNaiveMelPE(nn.Module):
-    def __init__(self, input_channels, out_dims, hidden_dims = 512, n_layers = 6, n_heads = 8, f0_max = 1975.5, f0_min = 32.70, use_fa_norm = False, conv_only = False, conv_dropout = 0, atten_dropout = 0, use_harmonic_emb = False):
-        super().__init__()
-        self.input_channels = input_channels
-        self.out_dims = out_dims
-        self.hidden_dims = hidden_dims
-        self.n_layers = n_layers
-        self.n_heads = n_heads
-        self.f0_max = f0_max
-        self.f0_min = f0_min
-        self.use_fa_norm = use_fa_norm
-        self.residual_dropout = 0.1  
-        self.attention_dropout = 0.1  
-        self.harmonic_emb = nn.Embedding(9, hidden_dims) if use_harmonic_emb else None
-        self.input_stack = nn.Sequential(nn.Conv1d(input_channels, hidden_dims, 3, 1, 1), nn.GroupNorm(4, hidden_dims), nn.LeakyReLU(), nn.Conv1d(hidden_dims, hidden_dims, 3, 1, 1))
-        self.net = ConformerNaiveEncoder(num_layers=n_layers, num_heads=n_heads, dim_model=hidden_dims, use_norm=use_fa_norm, conv_only=conv_only, conv_dropout=conv_dropout, atten_dropout=atten_dropout)
-        self.norm = nn.LayerNorm(hidden_dims)
-        self.output_proj = weight_norm(nn.Linear(hidden_dims, out_dims))
-        self.cent_table_b = torch.linspace(self.f0_to_cent(torch.Tensor([f0_min]))[0], self.f0_to_cent(torch.Tensor([f0_max]))[0], out_dims).detach()
-        self.register_buffer("cent_table", self.cent_table_b)
-        self.gaussian_blurred_cent_mask_b = (1200 * torch.log2(torch.Tensor([self.f0_max / 10.])))[0].detach()
-        self.register_buffer("gaussian_blurred_cent_mask", self.gaussian_blurred_cent_mask_b)
-
-    def forward(self, x, _h_emb=None):
-        x = self.input_stack(x.transpose(-1, -2)).transpose(-1, -2)
-        if self.harmonic_emb is not None: x = x + self.harmonic_emb(torch.LongTensor([0]).to(x.device)) if _h_emb is None else x + self.harmonic_emb(torch.LongTensor([int(_h_emb)]).to(x.device))
-        return torch.sigmoid(self.output_proj(self.norm(self.net(x))))
-
-    @torch.no_grad()
-    def latent2cents_decoder(self, y, threshold = 0.05, mask = True):
-        B, N, _ = y.size()
-        ci = self.cent_table[None, None, :].expand(B, N, -1)
-        rtn = torch.sum(ci * y, dim=-1, keepdim=True) / torch.sum(y, dim=-1, keepdim=True)  
-
-        if mask:
-            confident = torch.max(y, dim=-1, keepdim=True)[0]
-            confident_mask = torch.ones_like(confident)
-            confident_mask[confident <= threshold] = float("-INF")
-            rtn = rtn * confident_mask
-
-        return rtn  
-
-    @torch.no_grad()
-    def latent2cents_local_decoder(self, y, threshold = 0.05, mask = True):
-        B, N, _ = y.size()
-        ci = self.cent_table[None, None, :].expand(B, N, -1)
-        confident, max_index = torch.max(y, dim=-1, keepdim=True)
-
-        local_argmax_index = torch.arange(0, 9).to(max_index.device) + (max_index - 4)
-        local_argmax_index[local_argmax_index < 0] = 0
-        local_argmax_index[local_argmax_index >= self.out_dims] = self.out_dims - 1
-
-        y_l = torch.gather(y, -1, local_argmax_index)
-        rtn = torch.sum(torch.gather(ci, -1, local_argmax_index) * y_l, dim=-1, keepdim=True) / torch.sum(y_l, dim=-1, keepdim=True) 
-
-        if mask:
-            confident_mask = torch.ones_like(confident)
-            confident_mask[confident <= threshold] = float("-INF")
-            rtn = rtn * confident_mask
-
-        return rtn  
-
-    @torch.no_grad()
-    def infer(self, mel, decoder = "local_argmax", threshold = 0.05):
-        latent = self.forward(mel)
-        if decoder == "argmax": cents = self.latent2cents_local_decoder
-        elif decoder == "local_argmax": cents = self.latent2cents_local_decoder
-
-        return self.cent_to_f0(cents(latent, threshold=threshold))  
-
-    @torch.no_grad()
-    def cent_to_f0(self, cent: torch.Tensor) -> torch.Tensor:
-        return 10 * 2 ** (cent / 1200)
-
-    @torch.no_grad()
-    def f0_to_cent(self, f0):
-        return 1200 * torch.log2(f0 / 10)
-
-class CFNEncoderLayer(nn.Module):
-    def __init__(self, dim_model, num_heads = 8, use_norm = False, conv_only = False, conv_dropout = 0, atten_dropout = 0):
-        super().__init__()
-        self.conformer = nn.Sequential(ConformerConvModule(dim_model), nn.Dropout(conv_dropout)) if conv_dropout > 0 else ConformerConvModule(dim_model)
-        self.norm = nn.LayerNorm(dim_model)
-        self.dropout = nn.Dropout(0.1)  
-        self.attn = SelfAttention(dim=dim_model, heads=num_heads, causal=False, use_norm=use_norm, dropout=atten_dropout) if not conv_only else None
-
-    def forward(self, x, mask=None):
-        if self.attn is not None: x = x + (self.attn(self.norm(x), mask=mask))
-        return x + (self.conformer(x)) 
-
-class Swish(nn.Module):
-    def forward(self, x):
-        return x * x.sigmoid()
-
-class Transpose(nn.Module):
-    def __init__(self, dims):
-        super().__init__()
-        assert len(dims) == 2, "dims == 2"
-        self.dims = dims
-
-    def forward(self, x):
-        return x.transpose(*self.dims)
-
-class GLU(nn.Module):
-    def __init__(self, dim):
-        super().__init__()
-        self.dim = dim
-
-    def forward(self, x):
-        out, gate = x.chunk(2, dim=self.dim)
-        return out * gate.sigmoid()
-
-class DepthWiseConv1d_LEGACY(nn.Module):
-    def __init__(self, chan_in, chan_out, kernel_size, padding):
-        super().__init__()
-        self.padding = padding
-        self.conv = nn.Conv1d(chan_in, chan_out, kernel_size, groups=chan_in)
-
-    def forward(self, x):
-        return self.conv(F.pad(x, self.padding))
-
-class DepthWiseConv1d(nn.Module):
-    def __init__(self, chan_in, chan_out, kernel_size, padding, groups):
-        super().__init__()
-        self.conv = nn.Conv1d(chan_in, chan_out, kernel_size=kernel_size, padding=padding, groups=groups)
-
-    def forward(self, x):
-        return self.conv(x)
-
-class ConformerConvModule_LEGACY(nn.Module):
-    def __init__(self, dim, causal=False, expansion_factor=2, kernel_size=31, dropout=0.0):
-        super().__init__()
-        inner_dim = dim * expansion_factor
-        self.net = nn.Sequential(nn.LayerNorm(dim), Transpose((1, 2)), nn.Conv1d(dim, inner_dim * 2, 1), GLU(dim=1), DepthWiseConv1d_LEGACY(inner_dim, inner_dim, kernel_size=kernel_size, padding=(calc_same_padding(kernel_size) if not causal else (kernel_size - 1, 0))), Swish(), nn.Conv1d(inner_dim, dim, 1), Transpose((1, 2)), nn.Dropout(dropout))
-
-    def forward(self, x):
-        return self.net(x)
-
-class ConformerConvModule(nn.Module):
-    def __init__(self, dim, expansion_factor=2, kernel_size=31, dropout=0):
-        super().__init__()
-        inner_dim = dim * expansion_factor
-        self.net = nn.Sequential(nn.LayerNorm(dim), Transpose((1, 2)), nn.Conv1d(dim, inner_dim * 2, 1), nn.GLU(dim=1), DepthWiseConv1d(inner_dim, inner_dim, kernel_size=kernel_size, padding=calc_same_padding(kernel_size)[0], groups=inner_dim), nn.SiLU(), nn.Conv1d(inner_dim, dim, 1), Transpose((1, 2)), nn.Dropout(dropout))
-
-    def forward(self, x):
-        return self.net(x)
-
-class FastAttention(nn.Module):
-    def __init__(self, dim_heads, nb_features=None, ortho_scaling=0, causal=False, generalized_attention=False, kernel_fn=nn.ReLU(), qr_uniform_q=False, no_projection=False):
-        super().__init__()
-        nb_features = default(nb_features, int(dim_heads * math.log(dim_heads)))
-        self.dim_heads = dim_heads
-        self.nb_features = nb_features
-        self.ortho_scaling = ortho_scaling
-        self.create_projection = partial(gaussian_orthogonal_random_matrix, nb_rows=self.nb_features, nb_columns=dim_heads, scaling=ortho_scaling, qr_uniform_q=qr_uniform_q)
-        projection_matrix = self.create_projection()
-        self.register_buffer("projection_matrix", projection_matrix)
-        self.generalized_attention = generalized_attention
-        self.kernel_fn = kernel_fn
-        self.no_projection = no_projection
-        self.causal = causal
-
-    @torch.no_grad()
-    def redraw_projection_matrix(self):
-        projections = self.create_projection()
-        self.projection_matrix.copy_(projections)
-        del projections
-
-    def forward(self, q, k, v):
-        if self.no_projection: q, k = q.softmax(dim=-1), (torch.exp(k) if self.causal else k.softmax(dim=-2)) 
-        else:
-            create_kernel = partial(softmax_kernel, projection_matrix=self.projection_matrix, device=q.device)
-            q, k = create_kernel(q, is_query=True), create_kernel(k, is_query=False)
-
-        attn_fn = linear_attention if not self.causal else self.causal_linear_fn
-        return attn_fn(q, k, None) if v is None else attn_fn(q, k, v)
-
-class SelfAttention(nn.Module):
-    def __init__(self, dim, causal=False, heads=8, dim_head=64, local_heads=0, local_window_size=256, nb_features=None, feature_redraw_interval=1000, generalized_attention=False, kernel_fn=nn.ReLU(), qr_uniform_q=False, dropout=0.0, no_projection=False):
-        super().__init__()
-        assert dim % heads == 0
-        dim_head = default(dim_head, dim // heads)
-        inner_dim = dim_head * heads
-        self.fast_attention = FastAttention(dim_head, nb_features, causal=causal, generalized_attention=generalized_attention, kernel_fn=kernel_fn, qr_uniform_q=qr_uniform_q, no_projection=no_projection)
-        self.heads = heads
-        self.global_heads = heads - local_heads
-        self.local_attn = (LocalAttention(window_size=local_window_size, causal=causal, autopad=True, dropout=dropout, look_forward=int(not causal), rel_pos_emb_config=(dim_head, local_heads)) if local_heads > 0 else None)
-        self.to_q = nn.Linear(dim, inner_dim)
-        self.to_k = nn.Linear(dim, inner_dim)
-        self.to_v = nn.Linear(dim, inner_dim)
-        self.to_out = nn.Linear(inner_dim, dim)
-        self.dropout = nn.Dropout(dropout)
-
-    @torch.no_grad()
-    def redraw_projection_matrix(self):
-        self.fast_attention.redraw_projection_matrix()
-
-    def forward(self, x, context=None, mask=None, context_mask=None, name=None, inference=False, **kwargs):
-        _, _, _, h, gh = *x.shape, self.heads, self.global_heads
-        cross_attend = exists(context)
-        context = default(context, x)
-        context_mask = default(context_mask, mask) if not cross_attend else context_mask
-
-        q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=h), (self.to_q(x), self.to_k(context), self.to_v(context)))
-        (q, lq), (k, lk), (v, lv) = map(lambda t: (t[:, :gh], t[:, gh:]), (q, k, v))
-
-        attn_outs = []
-
-        if not empty(q):
-            if exists(context_mask): v.masked_fill_(~context_mask[:, None, :, None], 0.0)
-            if cross_attend: pass  
-            else: out = self.fast_attention(q, k, v)
-
-            attn_outs.append(out)
-
-        if not empty(lq):
-            assert (not cross_attend), "not cross_attend"
-
-            out = self.local_attn(lq, lk, lv, input_mask=mask)
-            attn_outs.append(out)
-
-        return self.dropout(self.to_out(rearrange(torch.cat(attn_outs, dim=1), "b h n d -> b n (h d)")))
-
-class HannWindow(torch.nn.Module):
-    def __init__(self, win_size):
-        super().__init__()
-        self.register_buffer('window', torch.hann_window(win_size), persistent=False)
-
-    def forward(self):
-        return self.window
-
-class FCPE_LEGACY(nn.Module):
-    def __init__(self, input_channel=128, out_dims=360, n_layers=12, n_chans=512, loss_mse_scale=10, loss_l2_regularization=False, loss_l2_regularization_scale=1, loss_grad1_mse=False, loss_grad1_mse_scale=1, f0_max=1975.5, f0_min=32.70, confidence=False, threshold=0.05, use_input_conv=True):
-        super().__init__()
-        self.loss_mse_scale = loss_mse_scale if (loss_mse_scale is not None) else 10
-        self.loss_l2_regularization = (loss_l2_regularization if (loss_l2_regularization is not None) else False)
-        self.loss_l2_regularization_scale = (loss_l2_regularization_scale if (loss_l2_regularization_scale is not None) else 1)
-        self.loss_grad1_mse = loss_grad1_mse if (loss_grad1_mse is not None) else False
-        self.loss_grad1_mse_scale = (loss_grad1_mse_scale if (loss_grad1_mse_scale is not None) else 1)
-        self.f0_max = f0_max if (f0_max is not None) else 1975.5
-        self.f0_min = f0_min if (f0_min is not None) else 32.70
-        self.confidence = confidence if (confidence is not None) else False
-        self.threshold = threshold if (threshold is not None) else 0.05
-        self.use_input_conv = use_input_conv if (use_input_conv is not None) else True
-        self.cent_table_b = torch.Tensor(np.linspace(self.f0_to_cent(torch.Tensor([f0_min]))[0], self.f0_to_cent(torch.Tensor([f0_max]))[0], out_dims))
-        self.register_buffer("cent_table", self.cent_table_b)
-        self.stack = nn.Sequential(nn.Conv1d(input_channel, n_chans, 3, 1, 1), nn.GroupNorm(4, n_chans), nn.LeakyReLU(), nn.Conv1d(n_chans, n_chans, 3, 1, 1))
-        self.decoder = PCmer(num_layers=n_layers, num_heads=8, dim_model=n_chans, dim_keys=n_chans, dim_values=n_chans, residual_dropout=0.1, attention_dropout=0.1)
-        self.norm = nn.LayerNorm(n_chans)
-        self.n_out = out_dims
-        self.dense_out = weight_norm(nn.Linear(n_chans, self.n_out))
-
-    def forward(self, mel, infer=True, gt_f0=None, return_hz_f0=False, cdecoder="local_argmax", output_interp_target_length=None):
-        if cdecoder == "argmax": self.cdecoder = self.cents_decoder
-        elif cdecoder == "local_argmax": self.cdecoder = self.cents_local_decoder
-
-        x = torch.sigmoid(self.dense_out(self.norm(self.decoder((self.stack(mel.transpose(1, 2)).transpose(1, 2) if self.use_input_conv else mel)))))
-
-        if not infer:
-            loss_all = self.loss_mse_scale * F.binary_cross_entropy(x, self.gaussian_blurred_cent(self.f0_to_cent(gt_f0)))
-            if self.loss_l2_regularization: loss_all = loss_all + l2_regularization(model=self, l2_alpha=self.loss_l2_regularization_scale)
-            x = loss_all
-        else:
-            x = self.cent_to_f0(self.cdecoder(x))
-            x = (1 + x / 700).log() if not return_hz_f0 else x
-
-        if output_interp_target_length is not None: 
-            x = F.interpolate(torch.where(x == 0, float("nan"), x).transpose(1, 2), size=int(output_interp_target_length), mode="linear").transpose(1, 2)
-            x = torch.where(x.isnan(), float(0.0), x)
-
-        return x
-
-    def cents_decoder(self, y, mask=True):
-        B, N, _ = y.size()
-        rtn = torch.sum(self.cent_table[None, None, :].expand(B, N, -1) * y, dim=-1, keepdim=True) / torch.sum(y, dim=-1, keepdim=True)
-
-        if mask:
-            confident = torch.max(y, dim=-1, keepdim=True)[0]
-            confident_mask = torch.ones_like(confident)
-            confident_mask[confident <= self.threshold] = float("-INF")
-            rtn = rtn * confident_mask
-
-        return (rtn, confident) if self.confidence else rtn
-
-    def cents_local_decoder(self, y, mask=True):
-        B, N, _ = y.size()
-
-        confident, max_index = torch.max(y, dim=-1, keepdim=True)
-        local_argmax_index = torch.clamp(torch.arange(0, 9).to(max_index.device) + (max_index - 4), 0, self.n_out - 1)
-        y_l = torch.gather(y, -1, local_argmax_index)
-        rtn = torch.sum(torch.gather(self.cent_table[None, None, :].expand(B, N, -1), -1, local_argmax_index) * y_l, dim=-1, keepdim=True) / torch.sum(y_l, dim=-1, keepdim=True)
-
-        if mask:
-            confident_mask = torch.ones_like(confident)
-            confident_mask[confident <= self.threshold] = float("-INF")
-            rtn = rtn * confident_mask
-
-        return (rtn, confident) if self.confidence else rtn
-
-    def cent_to_f0(self, cent):
-        return 10.0 * 2 ** (cent / 1200.0)
-
-    def f0_to_cent(self, f0):
-        return 1200.0 * torch.log2(f0 / 10.0)
-
-    def gaussian_blurred_cent(self, cents):
-        B, N, _ = cents.size()
-        return torch.exp(-torch.square(self.cent_table[None, None, :].expand(B, N, -1) - cents) / 1250) * (cents > 0.1) & (cents < (1200.0 * np.log2(self.f0_max / 10.0))).float()
-
-class InferCFNaiveMelPE(torch.nn.Module):
-    def __init__(self, args, state_dict):
-        super().__init__()
-        self.wav2mel = spawn_wav2mel(args, device="cpu")
-        self.model = CFNaiveMelPE(input_channels=catch_none_args_must(args.mel.num_mels, func_name="torchfcpe.tools.spawn_cf_naive_mel_pe", warning_str="args.mel.num_mels is None"), out_dims=catch_none_args_must(args.model.out_dims, func_name="torchfcpe.tools.spawn_cf_naive_mel_pe", warning_str="args.model.out_dims is None"), hidden_dims=catch_none_args_must(args.model.hidden_dims, func_name="torchfcpe.tools.spawn_cf_naive_mel_pe", warning_str="args.model.hidden_dims is None"), n_layers=catch_none_args_must(args.model.n_layers, func_name="torchfcpe.tools.spawn_cf_naive_mel_pe", warning_str="args.model.n_layers is None"), n_heads=catch_none_args_must(args.model.n_heads, func_name="torchfcpe.tools.spawn_cf_naive_mel_pe", warning_str="args.model.n_heads is None"), f0_max=catch_none_args_must(args.model.f0_max, func_name="torchfcpe.tools.spawn_cf_naive_mel_pe", warning_str="args.model.f0_max is None"), f0_min=catch_none_args_must(args.model.f0_min, func_name="torchfcpe.tools.spawn_cf_naive_mel_pe", warning_str="args.model.f0_min is None"), use_fa_norm=catch_none_args_must(args.model.use_fa_norm, func_name="torchfcpe.tools.spawn_cf_naive_mel_pe", warning_str="args.model.use_fa_norm is None"), conv_only=catch_none_args_opti(args.model.conv_only, default=False, func_name="torchfcpe.tools.spawn_cf_naive_mel_pe", warning_str="args.model.conv_only is None"), conv_dropout=catch_none_args_opti(args.model.conv_dropout, default=0.0, func_name="torchfcpe.tools.spawn_cf_naive_mel_pe", warning_str="args.model.conv_dropout is None"), atten_dropout=catch_none_args_opti(args.model.atten_dropout, default=0.0, func_name="torchfcpe.tools.spawn_cf_naive_mel_pe", warning_str="args.model.atten_dropout is None"), use_harmonic_emb=catch_none_args_opti(args.model.use_harmonic_emb, default=False, func_name="torchfcpe.tools.spawn_cf_naive_mel_pe", warning_str="args.model.use_harmonic_emb is None"))
-        self.model.load_state_dict(state_dict)
-        self.model.eval()
-        self.args_dict = dict(args)
-        self.register_buffer("tensor_device_marker", torch.tensor(1.0).float(), persistent=False)
-
-    def forward(self, wav, sr, decoder_mode = "local_argmax", threshold = 0.006, key_shifts = [0]):
-        with torch.no_grad():
-            mels = rearrange(torch.stack([self.wav2mel(wav.to(self.tensor_device_marker.device), sr, keyshift=keyshift) for keyshift in key_shifts], -1), "B T C K -> (B K) T C")
-            f0s = rearrange(self.model.infer(mels, decoder=decoder_mode, threshold=threshold), "(B K) T 1 -> B T (K 1)", K=len(key_shifts))
-
-        return f0s 
-
-    def infer(self, wav, sr, decoder_mode = "local_argmax", threshold = 0.006, f0_min = None, f0_max = None, interp_uv = False, output_interp_target_length = None, return_uv = False, test_time_augmentation = False, tta_uv_penalty = 12.0, tta_key_shifts = [0, -12, 12], tta_use_origin_uv=False):
-        if test_time_augmentation:
-            assert len(tta_key_shifts) > 0
-            flag = 0
-            if tta_use_origin_uv:
-                if 0 not in tta_key_shifts:
-                    flag = 1
-                    tta_key_shifts.append(0)
-
-            tta_key_shifts.sort(key=lambda x: (x if x >= 0 else -x / 2))
-            f0s = self.__call__(wav, sr, decoder_mode, threshold, tta_key_shifts)
-            f0 = ensemble_f0(f0s[:, :, flag:], tta_key_shifts[flag:], tta_uv_penalty)
-            f0_for_uv = f0s[:, :, [0]] if tta_use_origin_uv else f0
-        else:
-            f0 = self.__call__(wav, sr, decoder_mode, threshold)
-            f0_for_uv = f0
-
-        if f0_min is None: f0_min = self.args_dict["model"]["f0_min"]
-        uv = (f0_for_uv < f0_min).type(f0_for_uv.dtype)
-        f0 = f0 * (1 - uv)
-
-        if interp_uv: f0 = batch_interp_with_replacement_detach(uv.squeeze(-1).bool(), f0.squeeze(-1)).unsqueeze(-1)
-        if f0_max is not None: f0[f0 > f0_max] = f0_max
-        if output_interp_target_length is not None: 
-            f0 = F.interpolate(torch.where(f0 == 0, float("nan"), f0).transpose(1, 2), size=int(output_interp_target_length), mode="linear").transpose(1, 2)
-            f0 = torch.where(f0.isnan(), float(0.0), f0)
-
-        if return_uv: return f0, F.interpolate(uv.transpose(1, 2), size=int(output_interp_target_length), mode="nearest").transpose(1, 2)
-        else: return f0
-
-class FCPEInfer_LEGACY:
-    def __init__(self, model_path, device=None, dtype=torch.float32, providers=None, onnx=False, f0_min=50, f0_max=1100):
-        if device is None: device = "cuda" if torch.cuda.is_available() else "cpu"
-        self.device = device
-        self.dtype = dtype
-        self.onnx = onnx
-        self.f0_min = f0_min
-        self.f0_max = f0_max
-
-        if self.onnx:
-            sess_options = ort.SessionOptions()
-            sess_options.log_severity_level = 3
-            self.model = ort.InferenceSession(decrypt_model(model_path), sess_options=sess_options, providers=providers)
-        else:
-            ckpt = torch.load(model_path, map_location=torch.device(self.device))
-            self.args = DotDict(ckpt["config"])
-            model = FCPE_LEGACY(input_channel=self.args.model.input_channel, out_dims=self.args.model.out_dims, n_layers=self.args.model.n_layers, n_chans=self.args.model.n_chans, loss_mse_scale=self.args.loss.loss_mse_scale, loss_l2_regularization=self.args.loss.loss_l2_regularization, loss_l2_regularization_scale=self.args.loss.loss_l2_regularization_scale, loss_grad1_mse=self.args.loss.loss_grad1_mse, loss_grad1_mse_scale=self.args.loss.loss_grad1_mse_scale, f0_max=self.f0_max, f0_min=self.f0_min, confidence=self.args.model.confidence)
-            model.to(self.device).to(self.dtype)
-            model.load_state_dict(ckpt["model"])
-            model.eval()
-            self.model = model
-
-    @torch.no_grad()
-    def __call__(self, audio, sr, threshold=0.05, p_len=None):
-        if not self.onnx: self.model.threshold = threshold
-        self.wav2mel = Wav2Mel(device=self.device, dtype=self.dtype)
-
-        return (torch.as_tensor(self.model.run([self.model.get_outputs()[0].name], {self.model.get_inputs()[0].name: self.wav2mel(audio=audio[None, :], sample_rate=sr).to(self.dtype).detach().cpu().numpy(), self.model.get_inputs()[1].name: np.array(threshold, dtype=np.float32)})[0], dtype=self.dtype, device=self.device) if self.onnx else self.model(mel=self.wav2mel(audio=audio[None, :], sample_rate=sr).to(self.dtype), infer=True, return_hz_f0=True, output_interp_target_length=p_len))
-
-class FCPEInfer:
-    def __init__(self, model_path, device=None, dtype=torch.float32, providers=None, onnx=False, f0_min=50, f0_max=1100):
-        if device is None: device = "cuda" if torch.cuda.is_available() else "cpu"
-        self.device = device
-        self.dtype = dtype
-        self.onnx = onnx
-        self.f0_min = f0_min
-        self.f0_max = f0_max
-
-        if self.onnx:
-            sess_options = ort.SessionOptions()
-            sess_options.log_severity_level = 3
-            self.model = ort.InferenceSession(decrypt_model(model_path), sess_options=sess_options, providers=providers)
-        else:
-            ckpt = torch.load(model_path, map_location=torch.device(device))
-            ckpt["config_dict"]["model"]["conv_dropout"] = ckpt["config_dict"]["model"]["atten_dropout"] = 0.0
-            self.args = DotDict(ckpt["config_dict"])
-            model = InferCFNaiveMelPE(self.args, ckpt["model"])
-            model = model.to(device).to(self.dtype)
-            model.eval()
-            self.model = model
-
-    @torch.no_grad()
-    def __call__(self, audio, sr, threshold=0.05, p_len=None):
-        if self.onnx: self.wav2mel = Wav2Mel(device=self.device, dtype=self.dtype)
-        return (torch.as_tensor(self.model.run([self.model.get_outputs()[0].name], {self.model.get_inputs()[0].name: self.wav2mel(audio=audio[None, :], sample_rate=sr).to(self.dtype).detach().cpu().numpy(), self.model.get_inputs()[1].name: np.array(threshold, dtype=np.float32)})[0], dtype=self.dtype, device=self.device) if self.onnx else self.model.infer(audio[None, :], sr, threshold=threshold, f0_min=self.f0_min, f0_max=self.f0_max, output_interp_target_length=p_len))
-
-class MelModule(torch.nn.Module):
-    def __init__(self, sr, n_mels, n_fft, win_size, hop_length, fmin = None, fmax = None, clip_val = 1e-5, out_stft = False):
-        super().__init__()
-        if fmin is None: fmin = 0
-        if fmax is None: fmax = sr / 2
-        self.target_sr = sr
-        self.n_mels = n_mels
-        self.n_fft = n_fft
-        self.win_size = win_size
-        self.hop_length = hop_length
-        self.fmin = fmin
-        self.fmax = fmax
-        self.clip_val = clip_val
-        self.register_buffer('mel_basis', torch.tensor(librosa_mel_fn(sr=sr, n_fft=n_fft, n_mels=n_mels, fmin=fmin, fmax=fmax)).float(), persistent=False)
-        self.hann_window = torch.nn.ModuleDict()
-        self.out_stft = out_stft
-
-    @torch.no_grad()
-    def __call__(self, y, key_shift = 0, speed = 1, center = False, no_cache_window = False):
-        n_fft = self.n_fft
-        win_size = self.win_size
-        hop_length = self.hop_length
-        clip_val = self.clip_val
-        factor = 2 ** (key_shift / 12)
-        n_fft_new = int(np.round(n_fft * factor))
-        win_size_new = int(np.round(win_size * factor))
-        hop_length_new = int(np.round(hop_length * speed))
-
-        y = y.squeeze(-1)
-        if torch.min(y) < -1: print('[error with torchfcpe.mel_extractor.MelModule] min ', torch.min(y))
-        if torch.max(y) > 1: print('[error with torchfcpe.mel_extractor.MelModule] max ', torch.max(y))
-
-        key_shift_key = str(key_shift)
-        if not no_cache_window:
-            if key_shift_key in self.hann_window: hann_window = self.hann_window[key_shift_key]
-            else:
-                hann_window = HannWindow(win_size_new).to(self.mel_basis.device)
-                self.hann_window[key_shift_key] = hann_window
-
-            hann_window_tensor = hann_window()
-        else: hann_window_tensor = torch.hann_window(win_size_new).to(self.mel_basis.device)
-
-        pad_left = (win_size_new - hop_length_new) // 2
-        pad_right = max((win_size_new - hop_length_new + 1) // 2, win_size_new - y.size(-1) - pad_left)
-        mode = 'reflect' if pad_right < y.size(-1) else 'constant'
-        spec = torch.stft(F.pad(y.unsqueeze(1), (pad_left, pad_right), mode=mode).squeeze(1), n_fft_new, hop_length=hop_length_new, win_length=win_size_new, window=hann_window_tensor, center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=True)
-        spec = torch.sqrt(spec.real.pow(2) + spec.imag.pow(2) + 1e-9)
-
-        if key_shift != 0:
-            size = n_fft // 2 + 1
-            resize = spec.size(1)
-
-            if resize < size: spec = F.pad(spec, (0, 0, 0, size - resize))
-            spec = spec[:, :size, :] * win_size / win_size_new
-
-        spec = spec[:, :512, :] if self.out_stft else torch.matmul(self.mel_basis, spec)
-        return dynamic_range_compression_torch(spec, clip_val=clip_val).transpose(-1, -2)
-
-class Wav2MelModule(torch.nn.Module):
-    def __init__(self, sr, n_mels, n_fft, win_size, hop_length, fmin = None, fmax = None, clip_val = 1e-5, mel_type="default"):
-        super().__init__()
-        if fmin is None: fmin = 0
-        if fmax is None: fmax = sr / 2
-        self.sampling_rate = sr
-        self.n_mels = n_mels
-        self.n_fft = n_fft
-        self.win_size = win_size
-        self.hop_size = hop_length
-        self.fmin = fmin
-        self.fmax = fmax
-        self.clip_val = clip_val
-        self.register_buffer('tensor_device_marker', torch.tensor(1.0).float(), persistent=False)
-        self.resample_kernel = torch.nn.ModuleDict()
-        if mel_type == "default": self.mel_extractor = MelModule(sr, n_mels, n_fft, win_size, hop_length, fmin, fmax, clip_val, out_stft=False)
-        elif mel_type == "stft": self.mel_extractor = MelModule(sr, n_mels, n_fft, win_size, hop_length, fmin, fmax, clip_val, out_stft=True)
-        self.mel_type = mel_type
-
-    @torch.no_grad()
-    def __call__(self, audio, sample_rate, keyshift = 0, no_cache_window = False):
-        if sample_rate == self.sampling_rate: audio_res = audio
-        else:
-            key_str = str(sample_rate)
-            if key_str not in self.resample_kernel:
-                if len(self.resample_kernel) > 8: self.resample_kernel.clear()
-                self.resample_kernel[key_str] = Resample(sample_rate, self.sampling_rate, lowpass_filter_width=128).to(self.tensor_device_marker.device)
-
-            audio_res = self.resample_kernel[key_str](audio.squeeze(-1)).unsqueeze(-1)
-
-        mel = self.mel_extractor(audio_res, keyshift, no_cache_window=no_cache_window)
-        n_frames = int(audio.shape[1] // self.hop_size) + 1
-        if n_frames > int(mel.shape[1]): mel = torch.cat((mel, mel[:, -1:, :]), 1)
-        if n_frames < int(mel.shape[1]): mel = mel[:, :n_frames, :]
-
-        return mel 
-
-class Wav2Mel:
-    def __init__(self, device=None, dtype=torch.float32):
-        self.sample_rate = 16000
-        self.hop_size = 160
-        if device is None: device = "cuda" if torch.cuda.is_available() else "cpu"
-        self.device = device
-        self.dtype = dtype
-        self.stft = STFT(16000, 128, 1024, 1024, 160, 0, 8000)
-        self.resample_kernel = {}
-
-    def extract_nvstft(self, audio, keyshift=0, train=False):
-        return self.stft.get_mel(audio, keyshift=keyshift, train=train).transpose(1, 2)
-
-    def extract_mel(self, audio, sample_rate, keyshift=0, train=False):
-        audio = audio.to(self.dtype).to(self.device)
-        if sample_rate == self.sample_rate: audio_res = audio
-        else:
-            key_str = str(sample_rate)
-            if key_str not in self.resample_kernel: self.resample_kernel[key_str] = Resample(sample_rate, self.sample_rate, lowpass_filter_width=128)
-            self.resample_kernel[key_str] = (self.resample_kernel[key_str].to(self.dtype).to(self.device))
-            audio_res = self.resample_kernel[key_str](audio)
-
-        mel = self.extract_nvstft(audio_res, keyshift=keyshift, train=train) 
-        n_frames = int(audio.shape[1] // self.hop_size) + 1
-        mel = (torch.cat((mel, mel[:, -1:, :]), 1) if n_frames > int(mel.shape[1]) else mel)
-        return mel[:, :n_frames, :] if n_frames < int(mel.shape[1]) else mel
-
-    def __call__(self, audio, sample_rate, keyshift=0, train=False):
-        return self.extract_mel(audio, sample_rate, keyshift=keyshift, train=train)
-
-class DotDict(dict):
-    def __getattr__(*args):
-        val = dict.get(*args)
-        return DotDict(val) if type(val) is dict else val
-
-    __setattr__ = dict.__setitem__
-    __delattr__ = dict.__delitem__
-
-class FCPE:
-    def __init__(self, model_path, hop_length=512, f0_min=50, f0_max=1100, dtype=torch.float32, device=None, sample_rate=16000, threshold=0.05, providers=None, onnx=False, legacy=False):
-        self.model = FCPEInfer_LEGACY if legacy else FCPEInfer
-        self.fcpe = self.model(model_path, device=device, dtype=dtype, providers=providers, onnx=onnx, f0_min=f0_min, f0_max=f0_max)
-        self.hop_length = hop_length
-        self.device = device or ("cuda" if torch.cuda.is_available() else "cpu")
-        self.threshold = threshold
-        self.sample_rate = sample_rate
-        self.dtype = dtype
-        self.legacy = legacy
-
-    def repeat_expand(self, content, target_len, mode = "nearest"):
-        ndim = content.ndim
-        content = (content[None, None] if ndim == 1 else content[None] if ndim == 2 else content)
-
-        assert content.ndim == 3
-        is_np = isinstance(content, np.ndarray)
-
-        results = F.interpolate(torch.from_numpy(content) if is_np else content, size=target_len, mode=mode)
-        results = results.numpy() if is_np else results
-        return results[0, 0] if ndim == 1 else results[0] if ndim == 2 else results
-
-    def post_process(self, x, sample_rate, f0, pad_to):
-        f0 = (torch.from_numpy(f0).float().to(x.device) if isinstance(f0, np.ndarray) else f0)
-        f0 = self.repeat_expand(f0, pad_to) if pad_to is not None else f0
-
-        vuv_vector = torch.zeros_like(f0)
-        vuv_vector[f0 > 0.0] = 1.0
-        vuv_vector[f0 <= 0.0] = 0.0
-
-        nzindex = torch.nonzero(f0).squeeze()
-        f0 = torch.index_select(f0, dim=0, index=nzindex).cpu().numpy()
-        vuv_vector = F.interpolate(vuv_vector[None, None, :], size=pad_to)[0][0]
-
-        if f0.shape[0] <= 0: return np.zeros(pad_to), vuv_vector.cpu().numpy()
-        if f0.shape[0] == 1: return np.ones(pad_to) * f0[0], vuv_vector.cpu().numpy()
-        
-        return np.interp(np.arange(pad_to) * self.hop_length / sample_rate, self.hop_length / sample_rate * nzindex.cpu().numpy(), f0, left=f0[0], right=f0[-1]), vuv_vector.cpu().numpy()
-
-    def compute_f0(self, wav, p_len=None):
-        x = torch.FloatTensor(wav).to(self.dtype).to(self.device)
-        p_len = x.shape[0] // self.hop_length if p_len is None else p_len
-
-        f0 = self.fcpe(x, sr=self.sample_rate, threshold=self.threshold, p_len=p_len)
-        f0 = f0[:] if f0.dim() == 1 else f0[0, :, 0]
-
-        if torch.all(f0 == 0): return f0.cpu().numpy() if p_len is None else np.zeros(p_len), (f0.cpu().numpy() if p_len is None else np.zeros(p_len))
-        return self.post_process(x, self.sample_rate, f0, p_len)[0]

main/library/predictors/RMVPE.py

@@ -1,260 +0,0 @@
-import torch
-
-import numpy as np
-import torch.nn as nn
-import torch.nn.functional as F
-
-from librosa.filters import mel
-
-N_MELS, N_CLASS = 128, 360
-
-class ConvBlockRes(nn.Module):
-    def __init__(self, in_channels, out_channels, momentum=0.01):
-        super(ConvBlockRes, self).__init__()
-        self.conv = nn.Sequential(nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False), nn.BatchNorm2d(out_channels, momentum=momentum), nn.ReLU(), nn.Conv2d(in_channels=out_channels, out_channels=out_channels, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False), nn.BatchNorm2d(out_channels, momentum=momentum), nn.ReLU())
-
-        if in_channels != out_channels:
-            self.shortcut = nn.Conv2d(in_channels, out_channels, (1, 1))
-            self.is_shortcut = True
-        else: self.is_shortcut = False
-
-    def forward(self, x):
-        return self.conv(x) + self.shortcut(x) if self.is_shortcut else self.conv(x) + x
-
-class ResEncoderBlock(nn.Module):
-    def __init__(self, in_channels, out_channels, kernel_size, n_blocks=1, momentum=0.01):
-        super(ResEncoderBlock, self).__init__()
-        self.n_blocks = n_blocks
-        self.conv = nn.ModuleList()
-        self.conv.append(ConvBlockRes(in_channels, out_channels, momentum))
-
-        for _ in range(n_blocks - 1):
-            self.conv.append(ConvBlockRes(out_channels, out_channels, momentum))
-
-        self.kernel_size = kernel_size
-        if self.kernel_size is not None: self.pool = nn.AvgPool2d(kernel_size=kernel_size)
-
-    def forward(self, x):
-        for i in range(self.n_blocks):
-            x = self.conv[i](x)
-
-        if self.kernel_size is not None: return x, self.pool(x)
-        else: return x
-
-class Encoder(nn.Module):
-    def __init__(self, in_channels, in_size, n_encoders, kernel_size, n_blocks, out_channels=16, momentum=0.01):
-        super(Encoder, self).__init__()
-        self.n_encoders = n_encoders
-        self.bn = nn.BatchNorm2d(in_channels, momentum=momentum)
-        self.layers = nn.ModuleList()
-        self.latent_channels = []
-
-        for _ in range(self.n_encoders):
-            self.layers.append(ResEncoderBlock(in_channels, out_channels, kernel_size, n_blocks, momentum=momentum))
-            self.latent_channels.append([out_channels, in_size])
-            in_channels = out_channels
-            out_channels *= 2
-            in_size //= 2
-            
-        self.out_size = in_size
-        self.out_channel = out_channels
-
-    def forward(self, x):
-        concat_tensors = []
-        x = self.bn(x)
-
-        for i in range(self.n_encoders):
-            t, x = self.layers[i](x)
-            concat_tensors.append(t)
-
-        return x, concat_tensors
-
-class Intermediate(nn.Module):
-    def __init__(self, in_channels, out_channels, n_inters, n_blocks, momentum=0.01):
-        super(Intermediate, self).__init__()
-        self.n_inters = n_inters
-        self.layers = nn.ModuleList()
-        self.layers.append(ResEncoderBlock(in_channels, out_channels, None, n_blocks, momentum))
-
-        for _ in range(self.n_inters - 1):
-            self.layers.append(ResEncoderBlock(out_channels, out_channels, None, n_blocks, momentum))
-
-    def forward(self, x):
-        for i in range(self.n_inters):
-            x = self.layers[i](x)
-
-        return x
-
-class ResDecoderBlock(nn.Module):
-    def __init__(self, in_channels, out_channels, stride, n_blocks=1, momentum=0.01):
-        super(ResDecoderBlock, self).__init__()
-        out_padding = (0, 1) if stride == (1, 2) else (1, 1)
-        self.n_blocks = n_blocks
-        self.conv1 = nn.Sequential(nn.ConvTranspose2d(in_channels=in_channels, out_channels=out_channels, kernel_size=(3, 3), stride=stride, padding=(1, 1), output_padding=out_padding, bias=False), nn.BatchNorm2d(out_channels, momentum=momentum), nn.ReLU())
-        self.conv2 = nn.ModuleList()
-        self.conv2.append(ConvBlockRes(out_channels * 2, out_channels, momentum))
-
-        for _ in range(n_blocks - 1):
-            self.conv2.append(ConvBlockRes(out_channels, out_channels, momentum))
-
-    def forward(self, x, concat_tensor):
-        x = torch.cat((self.conv1(x), concat_tensor), dim=1)
-
-        for i in range(self.n_blocks):
-            x = self.conv2[i](x)
-
-        return x
-
-class Decoder(nn.Module):
-    def __init__(self, in_channels, n_decoders, stride, n_blocks, momentum=0.01):
-        super(Decoder, self).__init__()
-        self.layers = nn.ModuleList()
-        self.n_decoders = n_decoders
-
-        for _ in range(self.n_decoders):
-            out_channels = in_channels // 2
-            self.layers.append(ResDecoderBlock(in_channels, out_channels, stride, n_blocks, momentum))
-            in_channels = out_channels
-
-    def forward(self, x, concat_tensors):
-        for i in range(self.n_decoders):
-            x = self.layers[i](x, concat_tensors[-1 - i])
-
-        return x
-
-class DeepUnet(nn.Module):
-    def __init__(self, kernel_size, n_blocks, en_de_layers=5, inter_layers=4, in_channels=1, en_out_channels=16):
-        super(DeepUnet, self).__init__()
-        self.encoder = Encoder(in_channels, 128, en_de_layers, kernel_size, n_blocks, en_out_channels)
-        self.intermediate = Intermediate(self.encoder.out_channel // 2, self.encoder.out_channel, inter_layers, n_blocks)
-        self.decoder = Decoder(self.encoder.out_channel, en_de_layers, kernel_size, n_blocks)
-
-    def forward(self, x):
-        x, concat_tensors = self.encoder(x)
-        return self.decoder(self.intermediate(x), concat_tensors)
-
-class E2E(nn.Module):
-    def __init__(self, n_blocks, n_gru, kernel_size, en_de_layers=5, inter_layers=4, in_channels=1, en_out_channels=16):
-        super(E2E, self).__init__()
-        self.unet = DeepUnet(kernel_size, n_blocks, en_de_layers, inter_layers, in_channels, en_out_channels)
-        self.cnn = nn.Conv2d(en_out_channels, 3, (3, 3), padding=(1, 1))
-        self.fc = nn.Sequential(BiGRU(3 * 128, 256, n_gru), nn.Linear(512, N_CLASS), nn.Dropout(0.25), nn.Sigmoid()) if n_gru else nn.Sequential(nn.Linear(3 * N_MELS, N_CLASS), nn.Dropout(0.25), nn.Sigmoid())
-
-    def forward(self, mel):
-        return self.fc(self.cnn(self.unet(mel.transpose(-1, -2).unsqueeze(1))).transpose(1, 2).flatten(-2))
-
-class MelSpectrogram(torch.nn.Module):
-    def __init__(self, is_half, n_mel_channels, sample_rate, win_length, hop_length, n_fft=None, mel_fmin=0, mel_fmax=None, clamp=1e-5):
-        super().__init__()
-        n_fft = win_length if n_fft is None else n_fft
-        self.hann_window = {}
-        mel_basis = mel(sr=sample_rate, n_fft=n_fft, n_mels=n_mel_channels, fmin=mel_fmin, fmax=mel_fmax, htk=True)
-        mel_basis = torch.from_numpy(mel_basis).float()
-        self.register_buffer("mel_basis", mel_basis)
-        self.n_fft = win_length if n_fft is None else n_fft
-        self.hop_length = hop_length
-        self.win_length = win_length
-        self.sample_rate = sample_rate
-        self.n_mel_channels = n_mel_channels
-        self.clamp = clamp
-        self.is_half = is_half 
-
-    def forward(self, audio, keyshift=0, speed=1, center=True):
-        factor = 2 ** (keyshift / 12)
-        win_length_new = int(np.round(self.win_length * factor))
-        keyshift_key = str(keyshift) + "_" + str(audio.device)
-        if keyshift_key not in self.hann_window: self.hann_window[keyshift_key] = torch.hann_window(win_length_new).to(audio.device)
-
-        fft = torch.stft(audio, n_fft=int(np.round(self.n_fft * factor)), hop_length=int(np.round(self.hop_length * speed)), win_length=win_length_new, window=self.hann_window[keyshift_key], center=center, return_complex=True)
-        magnitude = torch.sqrt(fft.real.pow(2) + fft.imag.pow(2))
-
-        if keyshift != 0:
-            size = self.n_fft // 2 + 1
-            resize = magnitude.size(1)
-
-            if resize < size: magnitude = F.pad(magnitude, (0, 0, 0, size - resize))
-            magnitude = magnitude[:, :size, :] * self.win_length / win_length_new
-
-        mel_output = torch.matmul(self.mel_basis, magnitude)
-        if self.is_half: mel_output = mel_output.half()
-
-        return torch.log(torch.clamp(mel_output, min=self.clamp))
-
-class RMVPE:
-    def __init__(self, model_path, is_half, device=None, providers=None, onnx=False):
-        self.resample_kernel = {}
-        self.onnx = onnx
-
-        if self.onnx:
-            import onnxruntime as ort
-
-            sess_options = ort.SessionOptions()
-            sess_options.log_severity_level = 3
-
-            self.model = ort.InferenceSession(model_path, sess_options=sess_options, providers=providers)
-        else:
-            model = E2E(4, 1, (2, 2))
-            ckpt = torch.load(model_path, map_location="cpu")
-            model.load_state_dict(ckpt)
-            model.eval()
-            if is_half: model = model.half()
-            self.model = model.to(device)
-
-        self.resample_kernel = {}
-        self.is_half = is_half
-        self.device = device
-        self.mel_extractor = MelSpectrogram(is_half, N_MELS, 16000, 1024, 160, None, 30, 8000).to(device)
-        cents_mapping = 20 * np.arange(N_CLASS) + 1997.3794084376191
-        self.cents_mapping = np.pad(cents_mapping, (4, 4))
-
-    def mel2hidden(self, mel):
-        with torch.no_grad():
-            n_frames = mel.shape[-1]
-            mel = F.pad(mel, (0, 32 * ((n_frames - 1) // 32 + 1) - n_frames), mode="reflect")
-            hidden = self.model.run([self.model.get_outputs()[0].name], input_feed={self.model.get_inputs()[0].name: mel.cpu().numpy().astype(np.float32)})[0] if self.onnx else self.model(mel.half() if self.is_half else mel.float())
-            return hidden[:, :n_frames]
-
-    def decode(self, hidden, thred=0.03):
-        f0 = 10 * (2 ** (self.to_local_average_cents(hidden, thred=thred) / 1200))
-        f0[f0 == 10] = 0
-
-        return f0
-
-    def infer_from_audio(self, audio, thred=0.03):
-        hidden = self.mel2hidden(self.mel_extractor(torch.from_numpy(audio).float().to(self.device).unsqueeze(0), center=True))
-
-        return self.decode((hidden.squeeze(0).cpu().numpy().astype(np.float32) if self.is_half else hidden.squeeze(0).cpu().numpy()) if not self.onnx else hidden[0], thred=thred)
-    
-    def infer_from_audio_with_pitch(self, audio, thred=0.03, f0_min=50, f0_max=1100):
-        hidden = self.mel2hidden(self.mel_extractor(torch.from_numpy(audio).float().to(self.device).unsqueeze(0), center=True))
-
-        f0 = self.decode((hidden.squeeze(0).cpu().numpy().astype(np.float32) if self.is_half else hidden.squeeze(0).cpu().numpy()) if not self.onnx else hidden[0], thred=thred)
-        f0[(f0 < f0_min) | (f0 > f0_max)] = 0  
-
-        return f0
-
-    def to_local_average_cents(self, salience, thred=0.05):
-        center = np.argmax(salience, axis=1)
-        salience = np.pad(salience, ((0, 0), (4, 4)))
-        center += 4
-        todo_salience, todo_cents_mapping = [], []
-        starts = center - 4
-        ends = center + 5
-
-        for idx in range(salience.shape[0]):
-            todo_salience.append(salience[:, starts[idx] : ends[idx]][idx])
-            todo_cents_mapping.append(self.cents_mapping[starts[idx] : ends[idx]])
-
-        todo_salience = np.array(todo_salience)
-        devided = np.sum(todo_salience * np.array(todo_cents_mapping), 1) / np.sum(todo_salience, 1)
-        devided[np.max(salience, axis=1) <= thred] = 0
-
-        return devided
-
-class BiGRU(nn.Module):
-    def __init__(self, input_features, hidden_features, num_layers):
-        super(BiGRU, self).__init__()
-        self.gru = nn.GRU(input_features, hidden_features, num_layers=num_layers, batch_first=True, bidirectional=True)
-
-    def forward(self, x):
-        return self.gru(x)[0]

main/library/predictors/SWIPE.py

@@ -1,140 +0,0 @@
-import math
-
-import numpy as np
-
-from matplotlib import mlab
-from scipy import interpolate
-from decimal import Decimal, ROUND_HALF_UP
-
-def swipe(x, fs, f0_floor=50, f0_ceil=1100, frame_period=10, sTHR=0.3):
-    plim = np.array([f0_floor, f0_ceil])
-    t = np.arange(0, int(1000 * len(x) / fs / (frame_period) + 1)) * (frame_period / 1000)
-
-    log2pc = np.arange(np.log2(plim[0]) * 96, np.log2(plim[-1]) * 96)
-    log2pc *= (1 / 96)
-
-    pc = 2 ** log2pc
-    S = np.zeros((len(pc), len(t))) 
-
-    logWs = [round_matlab(elm) for elm in np.log2(4 * 2 * fs / plim)]
-    ws = 2 ** np.arange(logWs[0], logWs[1] - 1, -1) 
-    p0 = 4 * 2 * fs / ws 
-
-    d = 1 + log2pc - np.log2(4 * 2 * fs / ws[0])
-    fERBs = erbs2hz(np.arange(hz2erbs(pc[0] / 4), hz2erbs(fs / 2), 0.1))
-
-    for i in range(len(ws)):
-        dn = round_matlab(4 * fs / p0[i]) 
-        X, f, ti = mlab.specgram(x=np.r_[np.zeros(int(ws[i] / 2)), np.r_[x, np.zeros(int(dn + ws[i] / 2))]], NFFT=ws[i], Fs=fs, window=np.hanning(ws[i] + 2)[1:-1], noverlap=max(0, np.round(ws[i] - dn)), mode='complex')
-        ti = np.r_[0, ti[:-1]]
-        M = np.maximum(0, interpolate.interp1d(f, np.abs(X.T), kind='cubic')(fERBs)).T
-
-        if i == len(ws) - 1:
-            j = np.where(d - (i + 1) > -1)[0]
-            k = np.where(d[j] - (i + 1) < 0)[0]
-        elif i == 0:
-            j = np.where(d - (i + 1) < 1)[0]
-            k = np.where(d[j] - (i + 1) > 0)[0]
-        else:
-            j = np.where(np.abs(d - (i + 1)) < 1)[0]
-            k = np.arange(len(j))
-
-        Si = pitchStrengthAllCandidates(fERBs, np.sqrt(M), pc[j])
-        Si = interpolate.interp1d(ti, Si, bounds_error=False, fill_value='nan')(t) if Si.shape[1] > 1 else np.full((len(Si), len(t)), np.nan)
-
-        mu = np.ones(j.shape)
-        mu[k] = 1 - np.abs(d[j[k]] - i - 1)
-        S[j, :] = S[j, :] + np.tile(mu.reshape(-1, 1), (1, Si.shape[1])) * Si
-
-
-    p = np.full((S.shape[1], 1), np.nan)
-    s = np.full((S.shape[1], 1), np.nan)
-
-    for j in range(S.shape[1]):
-        s[j] = np.max(S[:, j])
-        i = np.argmax(S[:, j])
-
-        if s[j] < sTHR: continue
-
-        if i == 0: p[j] = pc[0]
-        elif i == len(pc) - 1: p[j] = pc[0]
-        else:
-            I = np.arange(i-1, i+2)
-            tc = 1 / pc[I]
-
-            ntc = (tc / tc[1] - 1) * 2 * np.pi
-            idx = np.isfinite(S[I, j])
-
-            c = np.zeros(len(ntc))
-            c += np.nan
-            
-            I_ = I[idx]
-
-            if len(I_) < 2: c[idx] = (S[I, j])[0] / ntc[0]
-            else: c[idx] = np.polyfit(ntc[idx], (S[I_, j]), 2)
-
-            pval = np.polyval(c, ((1 / (2 ** np.arange(np.log2(pc[I[0]]), np.log2(pc[I[2]]) + 1 / 12 / 64, 1 / 12 / 64))) / tc[1] - 1) * 2 * np.pi)
-            s[j] = np.max(pval)
-            p[j] = 2 ** (np.log2(pc[I[0]]) + (np.argmax(pval)) / 12 / 64)
-
-    p = p.flatten()
-    p[np.isnan(p)] = 0
-
-    return np.array(p, dtype=np.float32), np.array(t, dtype=np.float32)
-
-def round_matlab(n):
-    return int(Decimal(n).quantize(0, ROUND_HALF_UP))
-
-def pitchStrengthAllCandidates(f, L, pc):
-    den = np.sqrt(np.sum(L * L, axis=0))
-    den = np.where(den == 0, 2.220446049250313e-16, den)
-
-    L = L / den
-    S = np.zeros((len(pc), L.shape[1]))
-
-    for j in range(len(pc)):
-        S[j,:] = pitchStrengthOneCandidate(f, L, pc[j])
-
-    return S
-
-def pitchStrengthOneCandidate(f, L, pc):
-    k = np.zeros(len(f)) 
-    q = f / pc 
-
-    for i in ([1] + sieve(int(np.fix(f[-1] / pc - 0.75)))):
-        a = np.abs(q - i)
-        p = a < 0.25
-        k[p] = np.cos(2 * np.pi * q[p])
-
-        v = np.logical_and((0.25 < a), (a < 0.75))
-        k[v] = k[v] + np.cos(2 * np.pi * q[v]) / 2
-
-    k *= np.sqrt(1 / f)
-    k /= np.linalg.norm(k[k>0])
-
-    return k @ L
-
-def hz2erbs(hz):
-    return 21.4 * np.log10(1 + hz / 229)
-
-def erbs2hz(erbs):
-    return (10 ** (erbs / 21.4) - 1) * 229
-
-def sieve(n):
-    primes = list(range(2, n + 1))
-    num = 2
-
-    while num < math.sqrt(n):
-        i = num
-
-        while i <= n:
-            i += num
-
-            if i in primes: primes.remove(i)
-                
-        for j in primes:
-            if j > num:
-                num = j
-                break
-
-    return primes

main/library/predictors/WORLD_WRAPPER.py

@@ -1,90 +0,0 @@
-import os
-import torch
-import ctypes
-import platform
-
-import numpy as np
-
-
-
-class DioOption(ctypes.Structure):
-    _fields_ = [("F0Floor", ctypes.c_double), ("F0Ceil", ctypes.c_double), ("ChannelsInOctave", ctypes.c_double), ("FramePeriod", ctypes.c_double), ("Speed", ctypes.c_int), ("AllowedRange", ctypes.c_double)]
-
-class HarvestOption(ctypes.Structure):
-    _fields_ = [("F0Floor", ctypes.c_double), ("F0Ceil", ctypes.c_double), ("FramePeriod", ctypes.c_double)]
-
-class PYWORLD:
-    def __init__(self):
-        self.world_path = os.path.join("assets", "models", "predictors", "world")
-        os.makedirs(self.world_path, exist_ok=True)
-
-        model_type, suffix = (("world_64" if platform.architecture()[0] == "64bit" else "world_86"), ".dll") if platform.system() == "Windows" else ("world_linux", ".so")
-        self.world_file_path = os.path.join(self.world_path, f"{model_type}{suffix}")
-
-        if not os.path.exists(self.world_file_path):
-            model = torch.load(os.path.join("assets", "models", "predictors", "world.pth"), map_location="cpu")
-
-            with open(self.world_file_path, "wb") as w:
-                w.write(model[model_type])
-
-        self.world_dll = ctypes.CDLL(self.world_file_path)
-
-    def harvest(self, x, fs, f0_floor=50, f0_ceil=1100, frame_period=10):
-        self.world_dll.Harvest.argtypes = [ctypes.POINTER(ctypes.c_double), ctypes.c_int, ctypes.c_int, ctypes.POINTER(HarvestOption), ctypes.POINTER(ctypes.c_double), ctypes.POINTER(ctypes.c_double)]
-        self.world_dll.Harvest.restype = None 
-
-        self.world_dll.InitializeHarvestOption.argtypes = [ctypes.POINTER(HarvestOption)]
-        self.world_dll.InitializeHarvestOption.restype = None
-
-        self.world_dll.GetSamplesForHarvest.argtypes = [ctypes.c_int, ctypes.c_int, ctypes.c_double]
-        self.world_dll.GetSamplesForHarvest.restype = ctypes.c_int
-
-        option = HarvestOption()
-        self.world_dll.InitializeHarvestOption(ctypes.byref(option))
-
-        option.F0Floor = f0_floor
-        option.F0Ceil = f0_ceil
-        option.FramePeriod = frame_period
-
-        f0_length = self.world_dll.GetSamplesForHarvest(fs, len(x), option.FramePeriod)
-        f0 = (ctypes.c_double * f0_length)()
-        tpos = (ctypes.c_double * f0_length)()
-
-        self.world_dll.Harvest((ctypes.c_double * len(x))(*x), len(x), fs, ctypes.byref(option), tpos, f0)
-        return np.array(f0, dtype=np.float32), np.array(tpos, dtype=np.float32)
-
-    def dio(self, x, fs, f0_floor=50, f0_ceil=1100, channels_in_octave=2, frame_period=10, speed=1, allowed_range=0.1):
-        self.world_dll.Dio.argtypes = [ctypes.POINTER(ctypes.c_double), ctypes.c_int, ctypes.c_int, ctypes.POINTER(DioOption), ctypes.POINTER(ctypes.c_double), ctypes.POINTER(ctypes.c_double)]
-        self.world_dll.Dio.restype = None  
-
-        self.world_dll.InitializeDioOption.argtypes = [ctypes.POINTER(DioOption)]
-        self.world_dll.InitializeDioOption.restype = None
-
-        self.world_dll.GetSamplesForDIO.argtypes = [ctypes.c_int, ctypes.c_int, ctypes.c_double]
-        self.world_dll.GetSamplesForDIO.restype = ctypes.c_int
-
-        option = DioOption()
-        self.world_dll.InitializeDioOption(ctypes.byref(option))
-
-        option.F0Floor = f0_floor
-        option.F0Ceil = f0_ceil
-        option.ChannelsInOctave = channels_in_octave
-        option.FramePeriod = frame_period
-        option.Speed = speed
-        option.AllowedRange = allowed_range
-
-        f0_length = self.world_dll.GetSamplesForDIO(fs, len(x), option.FramePeriod)
-        f0 = (ctypes.c_double * f0_length)()
-        tpos = (ctypes.c_double * f0_length)()
-
-        self.world_dll.Dio((ctypes.c_double * len(x))(*x), len(x), fs, ctypes.byref(option), tpos, f0)
-        return np.array(f0, dtype=np.float32), np.array(tpos, dtype=np.float32)
-
-    def stonemask(self, x, fs, tpos, f0):
-        self.world_dll.StoneMask.argtypes = [ctypes.POINTER(ctypes.c_double), ctypes.c_int, ctypes.c_int, ctypes.POINTER(ctypes.c_double), ctypes.POINTER(ctypes.c_double), ctypes.c_int, ctypes.POINTER(ctypes.c_double)]
-        self.world_dll.StoneMask.restype = None 
-
-        out_f0 = (ctypes.c_double * len(f0))()
-        self.world_dll.StoneMask((ctypes.c_double * len(x))(*x), len(x), fs, (ctypes.c_double * len(tpos))(*tpos), (ctypes.c_double * len(f0))(*f0), len(f0), out_f0)
-
-        return np.array(out_f0, dtype=np.float32)

main/library/speaker_diarization/ECAPA_TDNN.py

@@ -1,280 +0,0 @@
-import math
-import torch  
-
-import torch.nn as nn
-import torch.nn.functional as F
-
-def length_to_mask(length, max_len=None, dtype=None, device=None):
-    assert len(length.shape) == 1
-
-    if max_len is None: max_len = length.max().long().item() 
-
-    mask = torch.arange(max_len, device=length.device, dtype=length.dtype).expand(len(length), max_len) < length.unsqueeze(1)
-
-    if dtype is None: dtype = length.dtype
-    if device is None: device = length.device
-
-    return torch.as_tensor(mask, dtype=dtype, device=device)
-
-def get_padding_elem(L_in, stride, kernel_size, dilation):
-    if stride > 1: padding = [math.floor(kernel_size / 2), math.floor(kernel_size / 2)]
-    else:
-        L_out = (math.floor((L_in - dilation * (kernel_size - 1) - 1) / stride) + 1)
-        padding = [math.floor((L_in - L_out) / 2), math.floor((L_in - L_out) / 2)]
-
-    return padding
-
-class _BatchNorm1d(nn.Module):
-    def __init__(self, input_shape=None, input_size=None, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True, combine_batch_time=False, skip_transpose=False):
-        super().__init__()
-        self.combine_batch_time = combine_batch_time
-        self.skip_transpose = skip_transpose
-
-        if input_size is None and skip_transpose: input_size = input_shape[1]
-        elif input_size is None: input_size = input_shape[-1]
-
-        self.norm = nn.BatchNorm1d(input_size, eps=eps, momentum=momentum, affine=affine, track_running_stats=track_running_stats)
-
-    def forward(self, x):
-        shape_or = x.shape
-
-        if self.combine_batch_time:x = x.reshape(shape_or[0] * shape_or[1], shape_or[2]) if x.ndim == 3 else x.reshape(shape_or[0] * shape_or[1], shape_or[3], shape_or[2])
-        elif not self.skip_transpose: x = x.transpose(-1, 1)
-
-        x_n = self.norm(x)
-
-        if self.combine_batch_time: x_n = x_n.reshape(shape_or)
-        elif not self.skip_transpose: x_n = x_n.transpose(1, -1)
-
-        return x_n
-
-class _Conv1d(nn.Module):
-    def __init__(self, out_channels, kernel_size, input_shape=None, in_channels=None, stride=1, dilation=1, padding="same", groups=1, bias=True, padding_mode="reflect", skip_transpose=False, weight_norm=False, conv_init=None, default_padding=0):
-        super().__init__()
-        self.kernel_size = kernel_size
-        self.stride = stride
-        self.dilation = dilation
-        self.padding = padding
-        self.padding_mode = padding_mode
-        self.unsqueeze = False
-        self.skip_transpose = skip_transpose
-
-        if input_shape is None and in_channels is None: raise ValueError
-        if in_channels is None: in_channels = self._check_input_shape(input_shape)
-
-        self.in_channels = in_channels
-        self.conv = nn.Conv1d(in_channels, out_channels, self.kernel_size, stride=self.stride, dilation=self.dilation, padding=default_padding, groups=groups, bias=bias)
-
-        if conv_init == "kaiming": nn.init.kaiming_normal_(self.conv.weight)
-        elif conv_init == "zero": nn.init.zeros_(self.conv.weight)
-        elif conv_init == "normal": nn.init.normal_(self.conv.weight, std=1e-6)
-
-        if weight_norm: self.conv = nn.utils.weight_norm(self.conv)
-
-    def forward(self, x):
-        if not self.skip_transpose: x = x.transpose(1, -1)
-        if self.unsqueeze: x = x.unsqueeze(1)
-
-        if self.padding == "same": x = self._manage_padding(x, self.kernel_size, self.dilation, self.stride)
-        elif self.padding == "causal": x = F.pad(x, ((self.kernel_size - 1) * self.dilation, 0))
-        elif self.padding == "valid": pass
-        else: raise ValueError
-
-        wx = self.conv(x)
-
-        if self.unsqueeze: wx = wx.squeeze(1)
-        if not self.skip_transpose: wx = wx.transpose(1, -1)
-
-        return wx
-
-    def _manage_padding(self, x, kernel_size, dilation, stride):
-        return F.pad(x, get_padding_elem(self.in_channels, stride, kernel_size, dilation), mode=self.padding_mode)
-
-    def _check_input_shape(self, shape):
-        if len(shape) == 2:
-            self.unsqueeze = True
-            in_channels = 1
-        elif self.skip_transpose: in_channels = shape[1]
-        elif len(shape) == 3: in_channels = shape[2]
-        else: raise ValueError
-
-        if not self.padding == "valid" and self.kernel_size % 2 == 0: raise ValueError
-        return in_channels
-
-    def remove_weight_norm(self):
-        self.conv = nn.utils.remove_weight_norm(self.conv)
-
-class Linear(torch.nn.Module):
-    def __init__(self, n_neurons, input_shape=None, input_size=None, bias=True, max_norm=None, combine_dims=False):
-        super().__init__()
-        self.max_norm = max_norm
-        self.combine_dims = combine_dims
-
-        if input_shape is None and input_size is None: raise ValueError
-        if input_size is None:
-            input_size = input_shape[-1]
-            if len(input_shape) == 4 and self.combine_dims: input_size = input_shape[2] * input_shape[3]
-
-        self.w = nn.Linear(input_size, n_neurons, bias=bias)
-
-    def forward(self, x):
-        if x.ndim == 4 and self.combine_dims: x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3])
-        if self.max_norm is not None: self.w.weight.data = torch.renorm(self.w.weight.data, p=2, dim=0, maxnorm=self.max_norm)
-
-        return self.w(x)
-
-class Conv1d(_Conv1d):
-    def __init__(self, *args, **kwargs):
-        super().__init__(skip_transpose=True, *args, **kwargs)
-
-class BatchNorm1d(_BatchNorm1d):
-    def __init__(self, *args, **kwargs):
-        super().__init__(skip_transpose=True, *args, **kwargs)
-
-class TDNNBlock(nn.Module):
-    def __init__(self, in_channels, out_channels, kernel_size, dilation, activation=nn.ReLU, groups=1, dropout=0.0):
-        super().__init__()
-        self.conv = Conv1d(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, dilation=dilation, groups=groups)
-        self.activation = activation()
-        self.norm = BatchNorm1d(input_size=out_channels)
-        self.dropout = nn.Dropout1d(p=dropout)
-
-    def forward(self, x):
-        return self.dropout(self.norm(self.activation(self.conv(x))))
-
-class Res2NetBlock(torch.nn.Module):
-    def __init__(self, in_channels, out_channels, scale=8, kernel_size=3, dilation=1, dropout=0.0):
-        super().__init__()
-        assert in_channels % scale == 0
-        assert out_channels % scale == 0
-        in_channel = in_channels // scale
-        hidden_channel = out_channels // scale
-        self.blocks = nn.ModuleList([TDNNBlock(in_channel, hidden_channel, kernel_size=kernel_size, dilation=dilation, dropout=dropout) for _ in range(scale - 1)])
-        self.scale = scale
-
-    def forward(self, x):
-        y = []
-
-        for i, x_i in enumerate(torch.chunk(x, self.scale, dim=1)):
-            if i == 0: y_i = x_i
-            elif i == 1: y_i = self.blocks[i - 1](x_i)
-            else: y_i = self.blocks[i - 1](x_i + y_i)
-
-            y.append(y_i)
-
-        return torch.cat(y, dim=1)
-
-class SEBlock(nn.Module):
-    def __init__(self, in_channels, se_channels, out_channels):
-        super().__init__()
-
-        self.conv1 = Conv1d(in_channels=in_channels, out_channels=se_channels, kernel_size=1)
-        self.relu = torch.nn.ReLU(inplace=True)
-        self.conv2 = Conv1d(in_channels=se_channels, out_channels=out_channels, kernel_size=1)
-        self.sigmoid = torch.nn.Sigmoid()
-
-    def forward(self, x, lengths=None):
-        L = x.shape[-1]
-
-        if lengths is not None:
-            mask = length_to_mask(lengths * L, max_len=L, device=x.device).unsqueeze(1)
-            s = (x * mask).sum(dim=2, keepdim=True) / mask.sum(dim=2, keepdim=True)
-        else: s = x.mean(dim=2, keepdim=True)
-
-        return self.sigmoid(self.conv2(self.relu(self.conv1(s)))) * x
-
-class AttentiveStatisticsPooling(nn.Module):
-    def __init__(self, channels, attention_channels=128, global_context=True):
-        super().__init__()
-        self.eps = 1e-12
-        self.global_context = global_context
-        self.tdnn = TDNNBlock(channels * 3, attention_channels, 1, 1) if global_context else TDNNBlock(channels, attention_channels, 1, 1)
-        self.tanh = nn.Tanh()
-        self.conv = Conv1d(in_channels=attention_channels, out_channels=channels, kernel_size=1)
-
-    def forward(self, x, lengths=None):
-        L = x.shape[-1]
-
-        def _compute_statistics(x, m, dim=2, eps=self.eps):
-            mean = (m * x).sum(dim)
-            return mean, torch.sqrt((m * (x - mean.unsqueeze(dim)).pow(2)).sum(dim).clamp(eps))
-
-        if lengths is None: lengths = torch.ones(x.shape[0], device=x.device)
-        mask = length_to_mask(lengths * L, max_len=L, device=x.device).unsqueeze(1)
-
-        if self.global_context:
-            mean, std = _compute_statistics(x, mask / mask.sum(dim=2, keepdim=True).float())
-            attn = torch.cat([x, mean.unsqueeze(2).repeat(1, 1, L), std.unsqueeze(2).repeat(1, 1, L)], dim=1)
-        else: attn = x
-
-        mean, std = _compute_statistics(x, F.softmax(self.conv(self.tanh(self.tdnn(attn))).masked_fill(mask == 0, float("-inf")), dim=2))
-        return torch.cat((mean, std), dim=1).unsqueeze(2)
-
-class SERes2NetBlock(nn.Module):
-    def __init__(self, in_channels, out_channels, res2net_scale=8, se_channels=128, kernel_size=1, dilation=1, activation=torch.nn.ReLU, groups=1, dropout=0.0):
-        super().__init__()
-        self.out_channels = out_channels
-        self.tdnn1 = TDNNBlock(in_channels, out_channels, kernel_size=1, dilation=1, activation=activation, groups=groups, dropout=dropout)
-        self.res2net_block = Res2NetBlock(out_channels, out_channels, res2net_scale, kernel_size, dilation)
-        self.tdnn2 = TDNNBlock(out_channels, out_channels, kernel_size=1, dilation=1, activation=activation, groups=groups, dropout=dropout)
-        self.se_block = SEBlock(out_channels, se_channels, out_channels)
-
-        self.shortcut = None
-        if in_channels != out_channels: self.shortcut = Conv1d(in_channels=in_channels, out_channels=out_channels, kernel_size=1)
-
-    def forward(self, x, lengths=None):
-        residual = x
-        if self.shortcut: residual = self.shortcut(x)
-
-        return self.se_block(self.tdnn2(self.res2net_block(self.tdnn1(x))), lengths) + residual
-
-class ECAPA_TDNN(torch.nn.Module):
-    def __init__(self, input_size, device="cpu", lin_neurons=192, activation=torch.nn.ReLU, channels=[512, 512, 512, 512, 1536], kernel_sizes=[5, 3, 3, 3, 1], dilations=[1, 2, 3, 4, 1], attention_channels=128, res2net_scale=8, se_channels=128, global_context=True, groups=[1, 1, 1, 1, 1], dropout=0.0):
-        super().__init__()
-        assert len(channels) == len(kernel_sizes)
-        assert len(channels) == len(dilations)
-
-        self.channels = channels
-        self.blocks = nn.ModuleList()
-
-        self.blocks.append(TDNNBlock(input_size, channels[0], kernel_sizes[0], dilations[0], activation, groups[0], dropout))
-
-        for i in range(1, len(channels) - 1):
-            self.blocks.append(SERes2NetBlock(channels[i - 1], channels[i], res2net_scale=res2net_scale, se_channels=se_channels, kernel_size=kernel_sizes[i], dilation=dilations[i], activation=activation, groups=groups[i], dropout=dropout))
-
-        self.mfa = TDNNBlock(channels[-2] * (len(channels) - 2), channels[-1], kernel_sizes[-1], dilations[-1], activation, groups=groups[-1], dropout=dropout)
-        self.asp = AttentiveStatisticsPooling(channels[-1], attention_channels=attention_channels, global_context=global_context)
-        self.asp_bn = BatchNorm1d(input_size=channels[-1] * 2)
-        self.fc = Conv1d(in_channels=channels[-1] * 2, out_channels=lin_neurons, kernel_size=1)
-
-    def forward(self, x, lengths=None):
-        x = x.transpose(1, 2)
-
-        xl = []
-        for layer in self.blocks:
-            try:
-                x = layer(x, lengths=lengths)
-            except TypeError:
-                x = layer(x)
-
-            xl.append(x)
-
-        return self.fc(self.asp_bn(self.asp(self.mfa(torch.cat(xl[1:], dim=1)), lengths=lengths))).transpose(1, 2)
-
-class Classifier(torch.nn.Module):
-    def __init__(self, input_size, device="cpu", lin_blocks=0, lin_neurons=192, out_neurons=1211):
-        super().__init__()
-        self.blocks = nn.ModuleList()
-
-        for _ in range(lin_blocks):
-            self.blocks.extend([_BatchNorm1d(input_size=input_size), Linear(input_size=input_size, n_neurons=lin_neurons)])
-            input_size = lin_neurons
-
-        self.weight = nn.Parameter(torch.FloatTensor(out_neurons, input_size, device=device))
-        nn.init.xavier_uniform_(self.weight)
-
-    def forward(self, x):
-        for layer in self.blocks:
-            x = layer(x)
-
-        return F.linear(F.normalize(x.squeeze(1)), F.normalize(self.weight)).unsqueeze(1)

main/library/speaker_diarization/audio.py

@@ -1,170 +0,0 @@
-import os
-import math
-import random
-import torchaudio
-
-from io import IOBase
-from torch.nn.functional import pad
-
-def get_torchaudio_info(file, backend = None):
-    if not backend:
-        backends = (torchaudio.list_audio_backends())
-        backend = "soundfile" if "soundfile" in backends else backends[0]
-
-    info = torchaudio.info(file["audio"], backend=backend)
-    if isinstance(file["audio"], IOBase): file["audio"].seek(0)
-
-    return info
-
-class Audio:
-    @staticmethod
-    def power_normalize(waveform):
-        return waveform / (waveform.square().mean(dim=-1, keepdim=True).sqrt() + 1e-8)
-
-    @staticmethod
-    def validate_file(file):
-        if isinstance(file, (str, os.PathLike)): file = {"audio": str(file), "uri": os.path.splitext(os.path.basename(file))[0]}
-        elif isinstance(file, IOBase): return {"audio": file, "uri": "stream"}
-        else: raise ValueError
-
-        if "waveform" in file:
-            waveform = file["waveform"]
-            if len(waveform.shape) != 2 or waveform.shape[0] > waveform.shape[1]: raise ValueError
-
-            sample_rate: int = file.get("sample_rate", None)
-            if sample_rate is None: raise ValueError
-
-            file.setdefault("uri", "waveform")
-
-        elif "audio" in file:
-            if isinstance(file["audio"], IOBase): return file
-
-            path = os.path.abspath(file["audio"])
-            file.setdefault("uri", os.path.splitext(os.path.basename(path))[0])
-
-        else: raise ValueError
-
-        return file
-
-    def __init__(self, sample_rate: int = None, mono=None, backend: str = None):
-        super().__init__()
-        self.sample_rate = sample_rate
-        self.mono = mono
-
-        if not backend:
-            backends = (torchaudio.list_audio_backends())  
-            backend = "soundfile" if "soundfile" in backends else backends[0]
-
-        self.backend = backend
-
-    def downmix_and_resample(self, waveform, sample_rate):
-        num_channels = waveform.shape[0]
-
-        if num_channels > 1:
-            if self.mono == "random":
-                channel = random.randint(0, num_channels - 1)
-                waveform = waveform[channel : channel + 1]
-            elif self.mono == "downmix": waveform = waveform.mean(dim=0, keepdim=True)
-
-        if (self.sample_rate is not None) and (self.sample_rate != sample_rate):
-            waveform = torchaudio.functional.resample(waveform, sample_rate, self.sample_rate)
-            sample_rate = self.sample_rate
-
-        return waveform, sample_rate
-
-    def get_duration(self, file):
-        file = self.validate_file(file)
-
-        if "waveform" in file:
-            frames = len(file["waveform"].T)
-            sample_rate = file["sample_rate"]
-        else:
-            info = file["torchaudio.info"] if "torchaudio.info" in file else get_torchaudio_info(file, backend=self.backend)
-            frames = info.num_frames
-            sample_rate = info.sample_rate
-
-        return frames / sample_rate
-
-    def get_num_samples(self, duration, sample_rate = None):
-        sample_rate = sample_rate or self.sample_rate
-        if sample_rate is None: raise ValueError
-
-        return math.floor(duration * sample_rate)
-
-    def __call__(self, file):
-        file = self.validate_file(file)
-
-        if "waveform" in file:
-            waveform = file["waveform"]
-            sample_rate = file["sample_rate"]
-        elif "audio" in file:
-            waveform, sample_rate = torchaudio.load(file["audio"], backend=self.backend)
-            if isinstance(file["audio"], IOBase): file["audio"].seek(0)
-
-        channel = file.get("channel", None)
-        if channel is not None: waveform = waveform[channel : channel + 1]
-
-        return self.downmix_and_resample(waveform, sample_rate)
-
-    def crop(self, file, segment, duration = None, mode="raise"):
-        file = self.validate_file(file)
-
-        if "waveform" in file:
-            waveform = file["waveform"]
-            frames = waveform.shape[1]
-            sample_rate = file["sample_rate"]
-        elif "torchaudio.info" in file:
-            info = file["torchaudio.info"]
-            frames = info.num_frames
-            sample_rate = info.sample_rate
-        else:
-            info = get_torchaudio_info(file, backend=self.backend)
-            frames = info.num_frames
-            sample_rate = info.sample_rate
-
-        channel = file.get("channel", None)
-        start_frame = math.floor(segment.start * sample_rate)
-
-        if duration:
-            num_frames = math.floor(duration * sample_rate)
-            end_frame = start_frame + num_frames
-        else:
-            end_frame = math.floor(segment.end * sample_rate)
-            num_frames = end_frame - start_frame
-
-        if mode == "raise":
-            if num_frames > frames: raise ValueError
-
-            if end_frame > frames + math.ceil(0.001 * sample_rate): raise ValueError
-            else:
-                end_frame = min(end_frame, frames)
-                start_frame = end_frame - num_frames
-
-            if start_frame < 0: raise ValueError
-        elif mode == "pad":
-            pad_start = -min(0, start_frame)
-            pad_end = max(end_frame, frames) - frames
-
-            start_frame = max(0, start_frame)
-            end_frame = min(end_frame, frames)
-
-            num_frames = end_frame - start_frame
-
-        if "waveform" in file: data = file["waveform"][:, start_frame:end_frame]
-        else:
-            try:
-                data, _ = torchaudio.load(file["audio"], frame_offset=start_frame, num_frames=num_frames, backend=self.backend)
-                if isinstance(file["audio"], IOBase): file["audio"].seek(0)
-            except RuntimeError:
-                if isinstance(file["audio"], IOBase): raise RuntimeError
-
-                waveform, sample_rate = self.__call__(file)
-                data = waveform[:, start_frame:end_frame]
-
-                file["waveform"] = waveform
-                file["sample_rate"] = sample_rate
-
-        if channel is not None: data = data[channel : channel + 1, :]
-        if mode == "pad": data = pad(data, (pad_start, pad_end))
-
-        return self.downmix_and_resample(data, sample_rate)

main/library/speaker_diarization/embedding.py

@@ -1,90 +0,0 @@
-import os
-import sys
-import torch
-
-import numpy as np
-import torch.nn.functional as F
-
-from functools import cached_property
-from torch.nn.utils.rnn import pad_sequence
-
-sys.path.append(os.getcwd())
-
-from main.library.speaker_diarization.speechbrain import EncoderClassifier
-
-class BaseInference:
-    pass
-
-class SpeechBrainPretrainedSpeakerEmbedding(BaseInference):
-    def __init__(self, embedding = "assets/models/speaker_diarization/models/speechbrain", device = None):
-        super().__init__()
-
-        self.embedding = embedding
-        self.device = device or torch.device("cpu")
-        self.classifier_ = EncoderClassifier.from_hparams(source=self.embedding, run_opts={"device": self.device})
-
-    def to(self, device):
-        if not isinstance(device, torch.device): raise TypeError
-
-        self.classifier_ = EncoderClassifier.from_hparams(source=self.embedding, run_opts={"device": device})
-        self.device = device
-        return self
-
-    @cached_property
-    def sample_rate(self):
-        return self.classifier_.audio_normalizer.sample_rate
-
-    @cached_property
-    def dimension(self):
-        *_, dimension = self.classifier_.encode_batch(torch.rand(1, 16000).to(self.device)).shape
-        return dimension
-
-    @cached_property
-    def metric(self):
-        return "cosine"
-
-    @cached_property
-    def min_num_samples(self):
-        with torch.inference_mode():
-            lower, upper = 2, round(0.5 * self.sample_rate)
-            middle = (lower + upper) // 2
-
-            while lower + 1 < upper:
-                try:
-                    _ = self.classifier_.encode_batch(torch.randn(1, middle).to(self.device))
-                    upper = middle
-                except RuntimeError:
-                    lower = middle
-
-                middle = (lower + upper) // 2
-
-        return upper
-
-    def __call__(self, waveforms, masks = None):
-        batch_size, num_channels, num_samples = waveforms.shape
-        assert num_channels == 1
-
-        waveforms = waveforms.squeeze(dim=1)
-
-        if masks is None:
-            signals = waveforms.squeeze(dim=1)
-            wav_lens = signals.shape[1] * torch.ones(batch_size)
-        else:
-            batch_size_masks, _ = masks.shape
-            assert batch_size == batch_size_masks
-
-            imasks = F.interpolate(masks.unsqueeze(dim=1), size=num_samples, mode="nearest").squeeze(dim=1) > 0.5
-            signals = pad_sequence([waveform[imask].contiguous() for waveform, imask in zip(waveforms, imasks)], batch_first=True)
-            wav_lens = imasks.sum(dim=1)
-
-        max_len = wav_lens.max()
-        if max_len < self.min_num_samples: return np.nan * np.zeros((batch_size, self.dimension))
-
-        too_short = wav_lens < self.min_num_samples
-        wav_lens = wav_lens / max_len
-        wav_lens[too_short] = 1.0
-
-        embeddings = (self.classifier_.encode_batch(signals, wav_lens=wav_lens).squeeze(dim=1).cpu().numpy())
-        embeddings[too_short.cpu().numpy()] = np.nan
-
-        return embeddings

main/library/speaker_diarization/encoder.py

@@ -1,250 +0,0 @@
-import os
-import sys
-import ast
-import torch
-import itertools
-import collections
-
-sys.path.append(os.getcwd())
-
-from main.library.speaker_diarization.speechbrain import if_main_process, ddp_barrier
-from main.library.speaker_diarization.features import register_checkpoint_hooks, mark_as_saver, mark_as_loader
-
-DEFAULT_UNK = "<unk>"
-DEFAULT_BOS = "<bos>"
-DEFAULT_EOS = "<eos>"
-DEFAULT_BLANK = "<blank>"
-
-@register_checkpoint_hooks
-class CategoricalEncoder:
-    VALUE_SEPARATOR = " => "
-    EXTRAS_SEPARATOR = "================\n"
-
-    def __init__(self, starting_index=0, **special_labels):
-        self.lab2ind = {}
-        self.ind2lab = {}
-        self.starting_index = starting_index
-        self.handle_special_labels(special_labels)
-
-    def handle_special_labels(self, special_labels):
-        if "unk_label" in special_labels: self.add_unk(special_labels["unk_label"])
-
-    def __len__(self):
-        return len(self.lab2ind)
-
-    @classmethod
-    def from_saved(cls, path):
-        obj = cls()
-        obj.load(path)
-        return obj
-
-    def update_from_iterable(self, iterable, sequence_input=False):
-        label_iterator = itertools.chain.from_iterable(iterable) if sequence_input else iter(iterable)
-        for label in label_iterator:
-            self.ensure_label(label)
-
-    def update_from_didataset(self, didataset, output_key, sequence_input=False):
-        with didataset.output_keys_as([output_key]):
-            self.update_from_iterable((data_point[output_key] for data_point in didataset), sequence_input=sequence_input)
-
-    def limited_labelset_from_iterable(self, iterable, sequence_input=False, n_most_common=None, min_count=1):
-        label_iterator = itertools.chain.from_iterable(iterable) if sequence_input else iter(iterable)
-        counts = collections.Counter(label_iterator)
-
-        for label, count in counts.most_common(n_most_common):
-            if count < min_count: break
-            self.add_label(label)
-
-        return counts
-
-    def load_or_create(self, path, from_iterables=[], from_didatasets=[], sequence_input=False, output_key=None, special_labels={}):
-        try:
-            if if_main_process():
-                if not self.load_if_possible(path):
-                    for iterable in from_iterables:
-                        self.update_from_iterable(iterable, sequence_input)
-
-                    for didataset in from_didatasets:
-                        if output_key is None: raise ValueError
-                        self.update_from_didataset(didataset, output_key, sequence_input)
-
-                    self.handle_special_labels(special_labels)
-                    self.save(path)
-        finally:
-            ddp_barrier()
-            self.load(path)
-
-    def add_label(self, label):
-        if label in self.lab2ind: raise KeyError
-        index = self._next_index()
-
-        self.lab2ind[label] = index
-        self.ind2lab[index] = label
-
-        return index
-
-    def ensure_label(self, label):
-        if label in self.lab2ind: return self.lab2ind[label]
-        else: return self.add_label(label)
-
-    def insert_label(self, label, index):
-        if label in self.lab2ind: raise KeyError
-        else: self.enforce_label(label, index)
-
-    def enforce_label(self, label, index):
-        index = int(index)
-
-        if label in self.lab2ind:
-            if index == self.lab2ind[label]: return
-            else: del self.ind2lab[self.lab2ind[label]]
-
-        if index in self.ind2lab:
-            saved_label = self.ind2lab[index]
-            moving_other = True
-        else: moving_other = False
-
-        self.lab2ind[label] = index
-        self.ind2lab[index] = label
-
-        if moving_other:
-            new_index = self._next_index()
-            self.lab2ind[saved_label] = new_index
-            self.ind2lab[new_index] = saved_label
-
-    def add_unk(self, unk_label=DEFAULT_UNK):
-        self.unk_label = unk_label
-        return self.add_label(unk_label)
-
-    def _next_index(self):
-        index = self.starting_index
-        while index in self.ind2lab:
-            index += 1
-
-        return index
-
-    def is_continuous(self):
-        indices = sorted(self.ind2lab.keys())
-        return self.starting_index in indices and all(j - i == 1 for i, j in zip(indices[:-1], indices[1:]))
-
-    def encode_label(self, label, allow_unk=True):
-        self._assert_len()
-
-        try:
-            return self.lab2ind[label]
-        except KeyError:
-            if hasattr(self, "unk_label") and allow_unk: return self.lab2ind[self.unk_label]
-            elif hasattr(self, "unk_label") and not allow_unk: raise KeyError
-            elif not hasattr(self, "unk_label") and allow_unk: raise KeyError
-            else: raise KeyError
-
-    def encode_label_torch(self, label, allow_unk=True):
-        return torch.LongTensor([self.encode_label(label, allow_unk)])
-
-    def encode_sequence(self, sequence, allow_unk=True):
-        self._assert_len()
-        return [self.encode_label(label, allow_unk) for label in sequence]
-
-    def encode_sequence_torch(self, sequence, allow_unk=True):
-        return torch.LongTensor([self.encode_label(label, allow_unk) for label in sequence])
-
-    def decode_torch(self, x):
-        self._assert_len()
-        decoded = []
-
-        if x.ndim == 1:  
-            for element in x:
-                decoded.append(self.ind2lab[int(element)])
-        else:
-            for subtensor in x:
-                decoded.append(self.decode_torch(subtensor))
-
-        return decoded
-
-    def decode_ndim(self, x):
-        self._assert_len()
-        try:
-            decoded = []
-            for subtensor in x:
-                decoded.append(self.decode_ndim(subtensor))
-
-            return decoded
-        except TypeError:  
-            return self.ind2lab[int(x)]
-
-    @mark_as_saver
-    def save(self, path):
-        self._save_literal(path, self.lab2ind, self._get_extras())
-
-    def load(self, path):
-        lab2ind, ind2lab, extras = self._load_literal(path)
-        self.lab2ind = lab2ind
-        self.ind2lab = ind2lab
-        self._set_extras(extras)
-
-    @mark_as_loader
-    def load_if_possible(self, path, end_of_epoch=False):
-        del end_of_epoch
-
-        try:
-            self.load(path)
-        except FileNotFoundError:
-            return False
-        except (ValueError, SyntaxError):
-            return False
-        
-        return True 
-
-    def expect_len(self, expected_len):
-        self.expected_len = expected_len
-
-    def ignore_len(self):
-        self.expected_len = None
-
-    def _assert_len(self):
-        if hasattr(self, "expected_len"):
-            if self.expected_len is None: return
-            if len(self) != self.expected_len: raise RuntimeError
-        else:
-            self.ignore_len()
-            return
-
-    def _get_extras(self):
-        extras = {"starting_index": self.starting_index}
-        if hasattr(self, "unk_label"): extras["unk_label"] = self.unk_label
-
-        return extras
-
-    def _set_extras(self, extras):
-        if "unk_label" in extras: self.unk_label = extras["unk_label"]
-        self.starting_index = extras["starting_index"]
-
-    @staticmethod
-    def _save_literal(path, lab2ind, extras):
-        with open(path, "w", encoding="utf-8") as f:
-            for label, ind in lab2ind.items():
-                f.write(repr(label) + CategoricalEncoder.VALUE_SEPARATOR + str(ind) + "\n")
-
-            f.write(CategoricalEncoder.EXTRAS_SEPARATOR)
-
-            for key, value in extras.items():
-                f.write(repr(key) + CategoricalEncoder.VALUE_SEPARATOR + repr(value) + "\n")
-
-            f.flush()
-
-    @staticmethod
-    def _load_literal(path):
-        lab2ind, ind2lab, extras = {}, {}, {}
-
-        with open(path, encoding="utf-8") as f:
-            for line in f:
-                if line == CategoricalEncoder.EXTRAS_SEPARATOR: break
-                literal, ind = line.strip().split(CategoricalEncoder.VALUE_SEPARATOR, maxsplit=1)
-                label = ast.literal_eval(literal)
-                lab2ind[label] = int(ind)
-                ind2lab[ind] = label
-
-            for line in f:
-                literal_key, literal_value = line.strip().split(CategoricalEncoder.VALUE_SEPARATOR, maxsplit=1)
-                extras[ast.literal_eval(literal_key)] = ast.literal_eval(literal_value)
-                
-        return lab2ind, ind2lab, extras

main/library/speaker_diarization/features.py

@@ -1,520 +0,0 @@
-import os
-import sys
-import math
-import torch
-import inspect
-import functools
-
-sys.path.append(os.getcwd())
-
-from main.library.speaker_diarization.speechbrain import MAIN_PROC_ONLY, is_distributed_initialized, main_process_only
-
-KEYS_MAPPING = {".mutihead_attn": ".multihead_attn",  ".convs_intermedite": ".convs_intermediate"}
-
-def map_old_state_dict_weights(state_dict, mapping):
-    for replacement_old, replacement_new in mapping.items():
-        for old_key in list(state_dict.keys()):
-            if replacement_old in old_key: state_dict[old_key.replace(replacement_old, replacement_new)] = state_dict.pop(old_key)
-
-    return state_dict
-
-def hook_on_loading_state_dict_checkpoint(state_dict):
-    return map_old_state_dict_weights(state_dict, KEYS_MAPPING)
-
-def torch_patched_state_dict_load(path, device="cpu"):
-    return hook_on_loading_state_dict_checkpoint(torch.load(path, map_location=device))
-
-@main_process_only
-def torch_save(obj, path):
-    state_dict = obj.state_dict()
-    torch.save(state_dict, path)
-
-def torch_recovery(obj, path, end_of_epoch):
-    del end_of_epoch  
-
-    state_dict = torch_patched_state_dict_load(path, "cpu")
-    try:
-        obj.load_state_dict(state_dict, strict=True)
-    except TypeError:
-        obj.load_state_dict(state_dict)
-
-def torch_parameter_transfer(obj, path):
-    incompatible_keys = obj.load_state_dict(torch_patched_state_dict_load(path, "cpu"), strict=False)
-
-    for missing_key in incompatible_keys.missing_keys:
-        pass
-    for unexpected_key in incompatible_keys.unexpected_keys:
-        pass
-
-WEAKREF_MARKER = "WEAKREF"
-
-def _cycliclrsaver(obj, path):
-    state_dict = obj.state_dict()
-    if state_dict.get("_scale_fn_ref") is not None: state_dict["_scale_fn_ref"] = WEAKREF_MARKER
-
-    torch.save(state_dict, path)
-
-def _cycliclrloader(obj, path, end_of_epoch):
-    del end_of_epoch  
-
-    try:
-        obj.load_state_dict(torch.load(path, map_location="cpu"), strict=True)
-    except TypeError:
-        obj.load_state_dict(torch.load(path, map_location="cpu"))
-
-DEFAULT_LOAD_HOOKS = {torch.nn.Module: torch_recovery, torch.optim.Optimizer: torch_recovery, torch.optim.lr_scheduler.ReduceLROnPlateau: torch_recovery, torch.cuda.amp.grad_scaler.GradScaler: torch_recovery}
-DEFAULT_SAVE_HOOKS = { torch.nn.Module: torch_save, torch.optim.Optimizer: torch_save, torch.optim.lr_scheduler.ReduceLROnPlateau: torch_save, torch.cuda.amp.grad_scaler.GradScaler: torch_save}
-DEFAULT_LOAD_HOOKS[torch.optim.lr_scheduler.LRScheduler] = torch_recovery
-DEFAULT_SAVE_HOOKS[torch.optim.lr_scheduler.LRScheduler] = torch_save
-DEFAULT_TRANSFER_HOOKS = {torch.nn.Module: torch_parameter_transfer}
-DEFAULT_SAVE_HOOKS[torch.optim.lr_scheduler.CyclicLR] = _cycliclrsaver
-DEFAULT_LOAD_HOOKS[torch.optim.lr_scheduler.CyclicLR] = _cycliclrloader
-
-def register_checkpoint_hooks(cls, save_on_main_only=True):
-    global DEFAULT_LOAD_HOOKS, DEFAULT_SAVE_HOOKS, DEFAULT_TRANSFER_HOOKS
-
-    for name, method in cls.__dict__.items():
-        if hasattr(method, "_speechbrain_saver"): DEFAULT_SAVE_HOOKS[cls] = main_process_only(method) if save_on_main_only else method
-        if hasattr(method, "_speechbrain_loader"): DEFAULT_LOAD_HOOKS[cls] = method
-        if hasattr(method, "_speechbrain_transfer"): DEFAULT_TRANSFER_HOOKS[cls] = method
-        
-    return cls
-
-def mark_as_saver(method):
-    sig = inspect.signature(method)
-
-    try:
-        sig.bind(object(), os.path.join("testpath"))
-    except TypeError:
-        raise TypeError
-    
-    method._speechbrain_saver = True
-    return method
-
-def mark_as_transfer(method):
-    sig = inspect.signature(method)
-    
-    try:
-        sig.bind(object(), os.path.join("testpath"))
-    except TypeError:
-        raise TypeError
-    
-    method._speechbrain_transfer = True
-    return method
-
-def mark_as_loader(method):
-    sig = inspect.signature(method)
-
-    try:
-        sig.bind(object(), os.path.join("testpath"), True)
-    except TypeError:
-        raise TypeError
-    
-    method._speechbrain_loader = True
-    return method
-
-def ddp_all_reduce(communication_object, reduce_op):
-    if MAIN_PROC_ONLY >= 1 or not is_distributed_initialized(): return communication_object
-    torch.distributed.all_reduce(communication_object, op=reduce_op)
-
-    return communication_object
-
-def fwd_default_precision(fwd = None, cast_inputs = torch.float32):
-    if fwd is None: return functools.partial(fwd_default_precision, cast_inputs=cast_inputs)
-
-    wrapped_fwd = torch.cuda.amp.custom_fwd(fwd, cast_inputs=cast_inputs)
-
-    @functools.wraps(fwd)
-    def wrapper(*args, force_allow_autocast = False, **kwargs):
-        return fwd(*args, **kwargs) if force_allow_autocast else wrapped_fwd(*args, **kwargs)
-
-    return wrapper
-
-def spectral_magnitude(stft, power = 1, log = False, eps = 1e-14):
-    spectr = stft.pow(2).sum(-1)
-
-    if power < 1: spectr = spectr + eps
-    spectr = spectr.pow(power)
-
-    if log: return torch.log(spectr + eps)
-    return spectr
-
-class Filterbank(torch.nn.Module):
-    def __init__(self, n_mels=40, log_mel=True, filter_shape="triangular", f_min=0, f_max=8000, n_fft=400, sample_rate=16000, power_spectrogram=2, amin=1e-10, ref_value=1.0, top_db=80.0, param_change_factor=1.0, param_rand_factor=0.0, freeze=True):
-        super().__init__()
-        self.n_mels = n_mels
-        self.log_mel = log_mel
-        self.filter_shape = filter_shape
-        self.f_min = f_min
-        self.f_max = f_max
-        self.n_fft = n_fft
-        self.sample_rate = sample_rate
-        self.power_spectrogram = power_spectrogram
-        self.amin = amin
-        self.ref_value = ref_value
-        self.top_db = top_db
-        self.freeze = freeze
-        self.n_stft = self.n_fft // 2 + 1
-        self.db_multiplier = math.log10(max(self.amin, self.ref_value))
-        self.device_inp = torch.device("cpu")
-        self.param_change_factor = param_change_factor
-        self.param_rand_factor = param_rand_factor
-        self.multiplier = 10 if self.power_spectrogram == 2 else 20
-
-        hz = self._to_hz(torch.linspace(self._to_mel(self.f_min), self._to_mel(self.f_max), self.n_mels + 2))
-
-        band = hz[1:] - hz[:-1]
-        self.band = band[:-1]
-        self.f_central = hz[1:-1]
-
-        if not self.freeze:
-            self.f_central = torch.nn.Parameter(self.f_central / (self.sample_rate * self.param_change_factor))
-            self.band = torch.nn.Parameter(self.band / (self.sample_rate * self.param_change_factor))
-
-        self.all_freqs_mat = torch.linspace(0, self.sample_rate // 2, self.n_stft).repeat(self.f_central.shape[0], 1)
-
-    def forward(self, spectrogram):
-        f_central_mat = self.f_central.repeat(self.all_freqs_mat.shape[1], 1).transpose(0, 1)
-        band_mat = self.band.repeat(self.all_freqs_mat.shape[1], 1).transpose(0, 1)
-
-        if not self.freeze:
-            f_central_mat = f_central_mat * (self.sample_rate * self.param_change_factor * self.param_change_factor)
-            band_mat = band_mat * (self.sample_rate * self.param_change_factor * self.param_change_factor)
-        elif self.param_rand_factor != 0 and self.training:
-            rand_change = (1.0 + torch.rand(2) * 2 * self.param_rand_factor - self.param_rand_factor)
-            f_central_mat = f_central_mat * rand_change[0]
-            band_mat = band_mat * rand_change[1]
-
-        fbank_matrix = self._create_fbank_matrix(f_central_mat, band_mat).to(spectrogram.device)
-        sp_shape = spectrogram.shape
-        if len(sp_shape) == 4: spectrogram = spectrogram.permute(0, 3, 1, 2).reshape(sp_shape[0] * sp_shape[3], sp_shape[1], sp_shape[2])
-
-        fbanks = torch.matmul(spectrogram, fbank_matrix)
-        if self.log_mel: fbanks = self._amplitude_to_DB(fbanks)
-
-        if len(sp_shape) == 4:
-            fb_shape = fbanks.shape
-            fbanks = fbanks.reshape(sp_shape[0], sp_shape[3], fb_shape[1], fb_shape[2]).permute(0, 2, 3, 1)
-
-        return fbanks
-
-    @staticmethod
-    def _to_mel(hz):
-        return 2595 * math.log10(1 + hz / 700)
-
-    @staticmethod
-    def _to_hz(mel):
-        return 700 * (10 ** (mel / 2595) - 1)
-
-    def _triangular_filters(self, all_freqs, f_central, band):
-        slope = (all_freqs - f_central) / band
-        return torch.max(torch.zeros(1, device=self.device_inp), torch.min(slope + 1.0, -slope + 1.0)).transpose(0, 1)
-
-    def _rectangular_filters(self, all_freqs, f_central, band):
-        left_side = right_size = all_freqs.ge(f_central - band)
-        right_size = all_freqs.le(f_central + band)
-
-        return (left_side * right_size).float().transpose(0, 1)
-
-    def _gaussian_filters(self, all_freqs, f_central, band, smooth_factor=torch.tensor(2)):
-        return torch.exp(-0.5 * ((all_freqs - f_central) / (band / smooth_factor)) ** 2).transpose(0, 1)
-
-    def _create_fbank_matrix(self, f_central_mat, band_mat):
-        if self.filter_shape == "triangular": fbank_matrix = self._triangular_filters(self.all_freqs_mat, f_central_mat, band_mat)
-        elif self.filter_shape == "rectangular": fbank_matrix = self._rectangular_filters(self.all_freqs_mat, f_central_mat, band_mat)
-        else: fbank_matrix = self._gaussian_filters(self.all_freqs_mat, f_central_mat, band_mat)
-
-        return fbank_matrix
-
-    def _amplitude_to_DB(self, x):
-        x_db = self.multiplier * torch.log10(torch.clamp(x, min=self.amin))
-        x_db -= self.multiplier * self.db_multiplier
-
-        return torch.max(x_db, (x_db.amax(dim=(-2, -1)) - self.top_db).view(x_db.shape[0], 1, 1))
-
-class ContextWindow(torch.nn.Module):
-    def __init__(self, left_frames=0, right_frames=0):
-        super().__init__()
-        self.left_frames = left_frames
-        self.right_frames = right_frames
-        self.context_len = self.left_frames + self.right_frames + 1
-        self.kernel_len = 2 * max(self.left_frames, self.right_frames) + 1
-        self.kernel = torch.eye(self.context_len, self.kernel_len)
-
-        if self.right_frames > self.left_frames: self.kernel = torch.roll(self.kernel, self.right_frames - self.left_frames, 1)
-        self.first_call = True
-
-    def forward(self, x):
-        x = x.transpose(1, 2)
-        if self.first_call:
-            self.first_call = False
-            self.kernel = (self.kernel.repeat(x.shape[1], 1, 1).view(x.shape[1] * self.context_len, self.kernel_len).unsqueeze(1))
-
-        or_shape = x.shape
-        if len(or_shape) == 4: x = x.reshape(or_shape[0] * or_shape[2], or_shape[1], or_shape[3])
-
-        cw_x = torch.nn.functional.conv1d(x, self.kernel.to(x.device), groups=x.shape[1], padding=max(self.left_frames, self.right_frames))
-        if len(or_shape) == 4: cw_x = cw_x.reshape(or_shape[0], cw_x.shape[1], or_shape[2], cw_x.shape[-1])
-
-        return cw_x.transpose(1, 2)
-
-class FilterProperties:
-    def __init__(self, window_size = 0, stride = 1, dilation = 1, causal = False):
-        self.window_size = window_size
-        self.stride = stride
-        self.dilation = dilation
-        self.causal = causal
-        
-    def __post_init__(self):
-        assert self.window_size > 0
-        assert self.stride > 0
-        assert (self.dilation > 0)
-
-    @staticmethod
-    def pointwise_filter():
-        return FilterProperties(window_size=1, stride=1)
-
-    def get_effective_size(self):
-        return 1 + ((self.window_size - 1) * self.dilation)
-
-    def get_convolution_padding(self):
-        if self.window_size % 2 == 0: raise ValueError
-        if self.causal: return self.get_effective_size() - 1
-
-        return (self.get_effective_size() - 1) // 2
-
-    def get_noncausal_equivalent(self):
-        if not self.causal: return self
-        return FilterProperties(window_size=(self.window_size - 1) * 2 + 1, stride=self.stride, dilation=self.dilation, causal=False)
-
-    def with_on_top(self, other, allow_approximate=True):
-        self_size = self.window_size
-
-        if other.window_size % 2 == 0:
-            if allow_approximate: other_size = other.window_size + 1
-            else: raise ValueError
-        else: other_size = other.window_size
-
-        if (self.causal or other.causal) and not (self.causal and other.causal):
-            if allow_approximate: return self.get_noncausal_equivalent().with_on_top(other.get_noncausal_equivalent())
-            else: raise ValueError
-
-        return FilterProperties(self_size + (self.stride * (other_size - 1)), self.stride * other.stride, self.dilation * other.dilation, self.causal)
-
-class STFT(torch.nn.Module):
-    def __init__(self, sample_rate, win_length=25, hop_length=10, n_fft=400, window_fn=torch.hamming_window, normalized_stft=False, center=True, pad_mode="constant", onesided=True):
-        super().__init__()
-        self.sample_rate = sample_rate
-        self.win_length = win_length
-        self.hop_length = hop_length
-        self.n_fft = n_fft
-        self.normalized_stft = normalized_stft
-        self.center = center
-        self.pad_mode = pad_mode
-        self.onesided = onesided
-        self.win_length = int(round((self.sample_rate / 1000.0) * self.win_length))
-        self.hop_length = int(round((self.sample_rate / 1000.0) * self.hop_length))
-        self.window = window_fn(self.win_length)
-
-    def forward(self, x):
-        or_shape = x.shape
-        if len(or_shape) == 3: x = x.transpose(1, 2).reshape(or_shape[0] * or_shape[2], or_shape[1])
-
-        stft = torch.view_as_real(torch.stft(x, self.n_fft, self.hop_length, self.win_length, self.window.to(x.device), self.center, self.pad_mode, self.normalized_stft, self.onesided, return_complex=True))
-        stft = stft.reshape(or_shape[0], or_shape[2], stft.shape[1], stft.shape[2], stft.shape[3]).permute(0, 3, 2, 4, 1) if len(or_shape) == 3 else stft.transpose(2, 1)
-
-        return stft
-
-    def get_filter_properties(self):
-        if not self.center: raise ValueError
-        return FilterProperties(window_size=self.win_length, stride=self.hop_length)
-
-class Deltas(torch.nn.Module):
-    def __init__(self, input_size, window_length=5):
-        super().__init__()
-        self.n = (window_length - 1) // 2
-        self.denom = self.n * (self.n + 1) * (2 * self.n + 1) / 3
-        self.register_buffer("kernel", torch.arange(-self.n, self.n + 1, dtype=torch.float32).repeat(input_size, 1, 1),)
-
-    def forward(self, x):
-        x = x.transpose(1, 2).transpose(2, -1)
-        or_shape = x.shape
-
-        if len(or_shape) == 4: x = x.reshape(or_shape[0] * or_shape[2], or_shape[1], or_shape[3])
-
-        x = torch.nn.functional.pad(x, (self.n, self.n), mode="replicate")
-        delta_coeff = (torch.nn.functional.conv1d(x, self.kernel.to(x.device), groups=x.shape[1]) / self.denom)
-
-        if len(or_shape) == 4: delta_coeff = delta_coeff.reshape(or_shape[0], or_shape[1], or_shape[2], or_shape[3])
-        return delta_coeff.transpose(1, -1).transpose(2, -1)
-
-class Fbank(torch.nn.Module):
-    def __init__(self, deltas=False, context=False, requires_grad=False, sample_rate=16000, f_min=0, f_max=None, n_fft=400, n_mels=40, filter_shape="triangular", param_change_factor=1.0, param_rand_factor=0.0, left_frames=5, right_frames=5, win_length=25, hop_length=10):
-        super().__init__()
-        self.deltas = deltas
-        self.context = context
-        self.requires_grad = requires_grad
-        if f_max is None: f_max = sample_rate / 2
-        self.compute_STFT = STFT(sample_rate=sample_rate,n_fft=n_fft,win_length=win_length,hop_length=hop_length)
-        self.compute_fbanks = Filterbank(sample_rate=sample_rate,n_fft=n_fft,n_mels=n_mels,f_min=f_min,f_max=f_max,freeze=not requires_grad,filter_shape=filter_shape,param_change_factor=param_change_factor,param_rand_factor=param_rand_factor)
-        self.compute_deltas = Deltas(input_size=n_mels)
-        self.context_window = ContextWindow(left_frames=left_frames, right_frames=right_frames)
-
-    @fwd_default_precision(cast_inputs=torch.float32)
-    def forward(self, wav):
-        fbanks = self.compute_fbanks(spectral_magnitude(self.compute_STFT(wav)))
-        if self.deltas:
-            delta1 = self.compute_deltas(fbanks)
-            fbanks = torch.cat([fbanks, delta1, self.compute_deltas(delta1)], dim=2)
-
-        if self.context: fbanks = self.context_window(fbanks)
-        return fbanks
-
-    def get_filter_properties(self):
-        return self.compute_STFT.get_filter_properties()
-
-@register_checkpoint_hooks
-class InputNormalization(torch.nn.Module):
-    def __init__(self, mean_norm=True, std_norm=True, norm_type="global", avg_factor=None, requires_grad=False, update_until_epoch=3):
-        super().__init__()
-        self.mean_norm = mean_norm
-        self.std_norm = std_norm
-        self.norm_type = norm_type
-        self.avg_factor = avg_factor
-        self.requires_grad = requires_grad
-        self.glob_mean = torch.tensor([0])
-        self.glob_std = torch.tensor([0])
-        self.spk_dict_mean = {}
-        self.spk_dict_std = {}
-        self.spk_dict_count = {}
-        self.weight = 1.0
-        self.count = 0
-        self.eps = 1e-10
-        self.update_until_epoch = update_until_epoch
-
-    def forward(self, x, lengths, spk_ids = torch.tensor([]), epoch=0):
-        N_batches = x.shape[0]
-        current_means, current_stds = [], []
-
-        if self.norm_type == "sentence" or self.norm_type == "speaker": out = torch.empty_like(x)
-
-        for snt_id in range(N_batches):
-            actual_size = torch.round(lengths[snt_id] * x.shape[1]).int()
-            current_mean, current_std = self._compute_current_stats(x[snt_id, 0:actual_size, ...])
-
-            current_means.append(current_mean)
-            current_stds.append(current_std)
-
-            if self.norm_type == "sentence": out[snt_id] = (x[snt_id] - current_mean.data) / current_std.data
-
-            if self.norm_type == "speaker":
-                spk_id = int(spk_ids[snt_id][0])
-
-                if self.training:
-                    if spk_id not in self.spk_dict_mean:
-                        self.spk_dict_mean[spk_id] = current_mean
-                        self.spk_dict_std[spk_id] = current_std
-                        self.spk_dict_count[spk_id] = 1
-                    else:
-                        self.spk_dict_count[spk_id] = (self.spk_dict_count[spk_id] + 1)
-                        self.weight = (1 / self.spk_dict_count[spk_id]) if self.avg_factor is None else self.avg_factor
-
-                        self.spk_dict_mean[spk_id] = (1 - self.weight) * self.spk_dict_mean[spk_id].to(current_mean) + self.weight * current_mean
-                        self.spk_dict_std[spk_id] = (1 - self.weight) * self.spk_dict_std[spk_id].to(current_std) + self.weight * current_std
-
-                        self.spk_dict_mean[spk_id].detach()
-                        self.spk_dict_std[spk_id].detach()
-
-                    speaker_mean = self.spk_dict_mean[spk_id].data
-                    speaker_std = self.spk_dict_std[spk_id].data
-                else:
-                    if spk_id in self.spk_dict_mean:
-                        speaker_mean = self.spk_dict_mean[spk_id].data
-                        speaker_std = self.spk_dict_std[spk_id].data
-                    else:
-                        speaker_mean = current_mean.data
-                        speaker_std = current_std.data
-
-                out[snt_id] = (x[snt_id] - speaker_mean) / speaker_std
-
-        if self.norm_type == "batch" or self.norm_type == "global":
-            current_mean = ddp_all_reduce(torch.mean(torch.stack(current_means), dim=0), torch.distributed.ReduceOp.AVG)
-            current_std = ddp_all_reduce(torch.mean(torch.stack(current_stds), dim=0), torch.distributed.ReduceOp.AVG)
-
-            if self.norm_type == "batch": out = (x - current_mean.data) / (current_std.data)
-
-            if self.norm_type == "global":
-                if self.training:
-                    if self.count == 0:
-                        self.glob_mean = current_mean
-                        self.glob_std = current_std
-                    elif epoch is None or epoch < self.update_until_epoch:
-                        self.weight = (1 / (self.count + 1)) if self.avg_factor is None else self.avg_factor
-                        self.glob_mean = (1 - self.weight) * self.glob_mean.to(current_mean) + self.weight * current_mean
-                        self.glob_std = (1 - self.weight) * self.glob_std.to(current_std) + self.weight * current_std
-
-                    self.glob_mean.detach()
-                    self.glob_std.detach()
-                    self.count = self.count + 1
-
-                out = (x - self.glob_mean.data.to(x)) / (self.glob_std.data.to(x))
-
-        return out
-
-    def _compute_current_stats(self, x):
-        current_std = torch.std(x, dim=0).detach().data if self.std_norm else torch.tensor([1.0], device=x.device)
-        return torch.mean(x, dim=0).detach().data if self.mean_norm else torch.tensor([0.0], device=x.device), torch.max(current_std, self.eps * torch.ones_like(current_std))
-
-    def _statistics_dict(self):
-        state = {}
-        state["count"] = self.count
-        state["glob_mean"] = self.glob_mean
-        state["glob_std"] = self.glob_std
-        state["spk_dict_mean"] = self.spk_dict_mean
-        state["spk_dict_std"] = self.spk_dict_std
-        state["spk_dict_count"] = self.spk_dict_count
-
-        return state
-
-    def _load_statistics_dict(self, state):
-        self.count = state["count"]
-
-        if isinstance(state["glob_mean"], int):
-            self.glob_mean = state["glob_mean"]
-            self.glob_std = state["glob_std"]
-        else:
-            self.glob_mean = state["glob_mean"]  
-            self.glob_std = state["glob_std"]  
-
-        self.spk_dict_mean = {}
-        for spk in state["spk_dict_mean"]:
-            self.spk_dict_mean[spk] = state["spk_dict_mean"][spk]
-
-        self.spk_dict_std = {}
-        for spk in state["spk_dict_std"]:
-            self.spk_dict_std[spk] = state["spk_dict_std"][spk] 
-
-        self.spk_dict_count = state["spk_dict_count"]
-        return state
-
-    def to(self, device):
-        self = super(InputNormalization, self).to(device)
-        self.glob_mean = self.glob_mean.to(device)
-        self.glob_std = self.glob_std.to(device)
-
-        for spk in self.spk_dict_mean:
-            self.spk_dict_mean[spk] = self.spk_dict_mean[spk].to(device)
-            self.spk_dict_std[spk] = self.spk_dict_std[spk].to(device)
-
-        return self
-
-    @mark_as_saver
-    def _save(self, path):
-        torch.save(self._statistics_dict(), path)
-
-    @mark_as_transfer
-    @mark_as_loader
-    def _load(self, path, end_of_epoch=False):
-        del end_of_epoch  
-        stats = torch.load(path, map_location="cpu")
-        self._load_statistics_dict(stats)

main/library/speaker_diarization/parameter_transfer.py

@@ -1,120 +0,0 @@
-import os
-import sys
-import inspect
-
-sys.path.append(os.getcwd())
-
-from main.library.speaker_diarization.speechbrain import fetch, run_on_main
-from main.library.speaker_diarization.features import DEFAULT_TRANSFER_HOOKS, DEFAULT_LOAD_HOOKS
-
-
-def get_default_hook(obj, default_hooks):
-    for cls in inspect.getmro(type(obj)):
-        if cls in default_hooks: return default_hooks[cls]
-        
-    return None
-
-class Pretrainer:
-    def __init__(self, loadables=None, paths=None, custom_hooks=None, conditions=None):
-        self.loadables = {}
-
-        if loadables is not None: self.add_loadables(loadables)
-        self.paths = {}
-
-        if paths is not None: self.add_paths(paths)
-        self.custom_hooks = {}
-
-        if custom_hooks is not None: self.add_custom_hooks(custom_hooks)
-        self.conditions = {}
-
-        if conditions is not None: self.add_conditions(conditions)
-        self.is_local = []
-
-    def add_loadables(self, loadables):
-        self.loadables.update(loadables)
-
-    def add_paths(self, paths):
-        self.paths.update(paths)
-
-    def add_custom_hooks(self, custom_hooks):
-        self.custom_hooks.update(custom_hooks)
-
-    def add_conditions(self, conditions):
-        self.conditions.update(conditions)
-
-    @staticmethod
-    def split_path(path):
-        def split(src):
-            if "/" in src: return src.rsplit("/", maxsplit=1)
-            else: return "./", src
-
-        return split(path)
-
-    def collect_files(self, default_source=None):
-        loadable_paths = {}
-        for name in self.loadables:
-            if not self.is_loadable(name): continue
-            save_filename = name + ".ckpt"
-
-            if name in self.paths: source, filename = self.split_path(self.paths[name])
-            elif default_source is not None:
-                filename = save_filename
-                source = default_source
-            else: raise ValueError
-
-            fetch_kwargs = {"filename": filename, "source": source}
-            path = None
-
-            def run_fetch(**kwargs):
-                nonlocal path
-
-                path = fetch(**kwargs)
-
-            run_on_main(run_fetch, kwargs=fetch_kwargs, post_func=run_fetch, post_kwargs=fetch_kwargs)
-
-            loadable_paths[name] = path
-            self.paths[name] = str(path)
-            self.is_local.append(name)
-
-        return loadable_paths
-
-    def is_loadable(self, name):
-        if name not in self.conditions: return True
-        condition = self.conditions[name]
-
-        if callable(condition): return condition()
-        else: return bool(condition)
-
-    def load_collected(self):
-        paramfiles = {}
-        for name in self.loadables:
-            if not self.is_loadable(name): continue
-
-            if name in self.is_local: paramfiles[name] = self.paths[name]
-            else: raise ValueError
-
-        self._call_load_hooks(paramfiles)
-
-    def _call_load_hooks(self, paramfiles):
-        for name, obj in self.loadables.items():
-            if not self.is_loadable(name): continue
-            loadpath = paramfiles[name]
-
-            if name in self.custom_hooks:
-                self.custom_hooks[name](obj, loadpath)
-                continue
-
-            default_hook = get_default_hook(obj, DEFAULT_TRANSFER_HOOKS)
-            
-            if default_hook is not None:
-                default_hook(obj, loadpath)
-                continue
-
-            default_hook = get_default_hook(obj, DEFAULT_LOAD_HOOKS)
-
-            if default_hook is not None:
-                end_of_epoch = False
-                default_hook(obj, loadpath, end_of_epoch)
-                continue
-
-            raise RuntimeError